diff --git "a/unet/dd/cache/NhwcConv_0-unetconv_inConv_meta.json" "b/unet/dd/cache/NhwcConv_0-unetconv_inConv_meta.json" --- "a/unet/dd/cache/NhwcConv_0-unetconv_inConv_meta.json" +++ "b/unet/dd/cache/NhwcConv_0-unetconv_inConv_meta.json" @@ -1,167745 +1,167745 @@ -{ - "dd_meta_major_version": 1, - "dd_meta_minor_version": 4, - "state_table_updates": [], - "op_list": [ - { - "name": "NhwcConv_0-/unet/conv_in/Conv", - "type": "SDConv", - "in_args": [ - "Transpose_34_out-sample.out0_1_0" - ], - "const_args": [ - "NhwcConv_0_weight_NHWC" - ], - "out_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "4" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "4" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_SDCastBf2Bfp_NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", - "type": "SDCastBf2Bfp", - "in_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0" - ], - "const_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts" - ], - "out_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_0", - "type": "SDGroupNorm_bfp", - "in_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" - ], - "const_args": [ - "GroupNorm_0_wts_6_0_0" - ], - "out_args": [ - "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/add_embedding/linear_1/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/Concat_3_output_0.out2_3_1" - ], - "const_args": [ - "unet.add_embedding.linear_1.weight_11_1_0" - ], - "out_args": [ - "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "2816" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2816", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid", - "type": "SDSilu", - "in_args": [ - "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1" - ], - "const_args": [ - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1" - ], - "out_args": [ - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "128" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/add_embedding/linear_2/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1" - ], - "const_args": [ - "unet.add_embedding.linear_2.weight_11_1_1" - ], - "out_args": [ - "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16751" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16752" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16775" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16776" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16807" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16808" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16831" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16832" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16863" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16864" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16887" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16888" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16911" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16912" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16935" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16936" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16959" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16960" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16983" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_16984" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17007" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17008" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17031" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17032" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17055" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17056" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17079" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17080" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17111" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17112" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17135" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17136" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17159" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17160" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17183" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17184" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17207" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17208" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17231" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17232" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17255" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17256" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17279" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17280" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17303" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17304" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17327" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17328" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17359" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17360" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17383" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17384" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17407" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17408" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17431" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17432" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17455" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17456" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17479" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17480" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17503" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17504" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17527" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17528" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17551" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17552" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17575" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17576" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/time_embedding/linear_1/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/time_proj/Concat_1_output_0.out2_3_0" - ], - "const_args": [ - "unet.time_embedding.linear_1.weight_11_1_2" - ], - "out_args": [ - "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid", - "type": "SDSilu", - "in_args": [ - "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0" - ], - "const_args": [ - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0" - ], - "out_args": [ - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "128" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/time_embedding/linear_2/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0" - ], - "const_args": [ - "unet.time_embedding.linear_2.weight_11_1_3" - ], - "out_args": [ - "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/Add", - "type": "SDAdd", - "in_args": [ - "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2", - "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3" - ], - "const_args": [], - "out_args": [ - "/unet/Add_output_0.out_17_1_0" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid", - "type": "SDSilu", - "in_args": [ - "/unet/Add_output_0.out_17_1_0" - ], - "const_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2" - ], - "out_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "128" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4" - ], - "out_args": [ - "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_1-/unet/down_blocks.0/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0" - ], - "const_args": [ - "NhwcConv_1_weight_NHWC" - ], - "out_args": [ - "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1", - "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_1", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1" - ], - "const_args": [ - "GroupNorm_1_wts_6_0_1" - ], - "out_args": [ - "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_2-/unet/down_blocks.0/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1" - ], - "const_args": [ - "NhwcConv_2_weight_NHWC" - ], - "out_args": [ - "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", - "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_2", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" - ], - "const_args": [ - "GroupNorm_2_wts_6_0_2" - ], - "out_args": [ - "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5" - ], - "out_args": [ - "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_3-/unet/down_blocks.0/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2" - ], - "const_args": [ - "NhwcConv_3_weight_NHWC" - ], - "out_args": [ - "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3", - "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_3", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3" - ], - "const_args": [ - "GroupNorm_3_wts_6_0_3" - ], - "out_args": [ - "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_4-/unet/down_blocks.0/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3" - ], - "const_args": [ - "NhwcConv_4_weight_NHWC" - ], - "out_args": [ - "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.0/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2", - "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_5-/unet/down_blocks.0/downsamplers.0/conv/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" - ], - "const_args": [ - "NhwcConv_5_weight_NHWC" - ], - "out_args": [ - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "2", - "2" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_4", - "type": "SDGroupNorm_bfp", - "in_args": [ - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" - ], - "const_args": [ - "GroupNorm_4_wts_6_0_4" - ], - "out_args": [ - "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6" - ], - "out_args": [ - "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_6-/unet/down_blocks.1/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4" - ], - "const_args": [ - "NhwcConv_6_weight_NHWC" - ], - "out_args": [ - "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7", - "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_5", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5" - ], - "const_args": [ - "GroupNorm_5_wts_6_0_5" - ], - "out_args": [ - "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7" - ], - "out_args": [ - "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8" - ], - "out_args": [ - "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9" - ], - "out_args": [ - "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10" - ], - "out_args": [ - "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11" - ], - "out_args": [ - "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17611" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17612" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17635" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17636" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17659" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17660" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17683" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17684" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17707" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17708" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17731" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17732" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17755" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17756" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17779" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17780" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17803" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17804" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17827" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17828" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17859" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17860" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17883" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17884" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17907" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17908" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17931" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17932" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17955" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17956" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17979" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_17980" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18003" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18004" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18027" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18028" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18051" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18052" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18075" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18076" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18107" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18108" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18131" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18132" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18155" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18156" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18179" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18180" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18203" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18204" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18227" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18228" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18251" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18252" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18275" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18276" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18299" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18300" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18323" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18324" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12" - ], - "out_args": [ - "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13" - ], - "out_args": [ - "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.2/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14" - ], - "out_args": [ - "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18356" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18357" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18380" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18381" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18412" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18413" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18436" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18437" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18468" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18469" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18492" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul", - "type": "SDGemm", - "in_args": [ - "encoder_hidden_states.out2_6_0" - ], - "const_args": [ - "onnx::MatMul_18493" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "77", - "2048" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "77", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2048", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15" - ], - "out_args": [ - "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16" - ], - "out_args": [ - "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.2/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17" - ], - "out_args": [ - "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.0/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18" - ], - "out_args": [ - "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.1/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19" - ], - "out_args": [ - "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.2/time_emb_proj/Gemm", - "type": "SDGemm", - "in_args": [ - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" - ], - "const_args": [ - "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20" - ], - "out_args": [ - "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_7-/unet/down_blocks.1/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5" - ], - "const_args": [ - "NhwcConv_7_weight_NHWC" - ], - "out_args": [ - "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_8-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" - ], - "const_args": [ - "NhwcConv_8_weight_NHWC" - ], - "out_args": [ - "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "1", - "1", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6", - "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_6", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" - ], - "const_args": [ - "GroupNorm_6_wts_6_1_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6" - ], - "const_args": [ - "onnx::MatMul_16738_11_0_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" - ], - "const_args": [ - "Attention_0_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" - ], - "const_args": [ - "Attention_0_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" - ], - "const_args": [ - "Attention_0_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_0", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0" - ], - "const_args": [ - "onnx::MatMul_16749_11_0_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235", - "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1" - ], - "const_args": [ - "onnx::MatMul_16750" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_SDCastBfp2Bf_/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_0", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_SDCastBf2Bfp_/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236" - ], - "const_args": [ - "onnx::MatMul_16760_11_0_2" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_0_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" - ], - "const_args": [ - "onnx::MatMul_16761_11_0_3_12_0_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_0_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" - ], - "const_args": [ - "onnx::MatMul_16761_11_0_3_12_0_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0" - ], - "const_args": [ - "onnx::MatMul_16762_11_0_4" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" - ], - "const_args": [ - "Attention_1_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" - ], - "const_args": [ - "Attention_1_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" - ], - "const_args": [ - "Attention_1_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_1", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1" - ], - "const_args": [ - "onnx::MatMul_16773_11_0_5" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4" - ], - "const_args": [ - "onnx::MatMul_16774" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_SDCastBfp2Bf_/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_1", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_SDCastBf2Bfp_/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244" - ], - "const_args": [ - "onnx::MatMul_16784_11_0_6" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" - ], - "const_args": [ - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_1_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" - ], - "const_args": [ - "onnx::MatMul_16785_11_0_7_12_1_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_1_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" - ], - "const_args": [ - "onnx::MatMul_16785_11_0_7_12_1_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1" - ], - "const_args": [ - "onnx::MatMul_16786_11_0_8" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12" - ], - "const_args": [ - "onnx::MatMul_16787_11_0_9" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248", - "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_7", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" - ], - "const_args": [ - "GroupNorm_7_wts_6_0_6" - ], - "out_args": [ - "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_9-/unet/down_blocks.1/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7" - ], - "const_args": [ - "NhwcConv_9_weight_NHWC" - ], - "out_args": [ - "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9", - "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_8", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14" - ], - "const_args": [ - "GroupNorm_8_wts_6_0_7" - ], - "out_args": [ - "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_10-/unet/down_blocks.1/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8" - ], - "const_args": [ - "NhwcConv_10_weight_NHWC" - ], - "out_args": [ - "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13", - "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_9", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" - ], - "const_args": [ - "GroupNorm_9_wts_6_1_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9" - ], - "const_args": [ - "onnx::MatMul_16794_11_0_10" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" - ], - "const_args": [ - "Attention_2_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" - ], - "const_args": [ - "Attention_2_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" - ], - "const_args": [ - "Attention_2_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_2", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2" - ], - "const_args": [ - "onnx::MatMul_16805_11_0_11" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253", - "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7" - ], - "const_args": [ - "onnx::MatMul_16806" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_SDCastBfp2Bf_/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_2", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_SDCastBf2Bfp_/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254" - ], - "const_args": [ - "onnx::MatMul_16816_11_0_12" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_2_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" - ], - "const_args": [ - "onnx::MatMul_16817_11_0_13_12_2_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_2_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" - ], - "const_args": [ - "onnx::MatMul_16817_11_0_13_12_2_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2" - ], - "const_args": [ - "onnx::MatMul_16818_11_0_14" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" - ], - "const_args": [ - "Attention_3_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" - ], - "const_args": [ - "Attention_3_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" - ], - "const_args": [ - "Attention_3_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_3", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3" - ], - "const_args": [ - "onnx::MatMul_16829_11_0_15" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10" - ], - "const_args": [ - "onnx::MatMul_16830" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_SDCastBfp2Bf_/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_3", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_SDCastBf2Bfp_/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7" - ], - "const_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262" - ], - "const_args": [ - "onnx::MatMul_16840_11_0_16" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" - ], - "const_args": [ - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_3_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" - ], - "const_args": [ - "onnx::MatMul_16841_11_0_17_12_3_0" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_3_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" - ], - "const_args": [ - "onnx::MatMul_16841_11_0_17_12_3_1" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3" - ], - "const_args": [ - "onnx::MatMul_16842_11_0_18" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21" - ], - "const_args": [ - "onnx::MatMul_16843_11_0_19" - ], - "out_args": [ - "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.1/attentions.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266", - "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_11-/unet/down_blocks.1/downsamplers.0/conv/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" - ], - "const_args": [ - "NhwcConv_11_weight_NHWC" - ], - "out_args": [ - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "2", - "2" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_10", - "type": "SDGroupNorm_bfp", - "in_args": [ - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" - ], - "const_args": [ - "GroupNorm_10_wts_6_0_8" - ], - "out_args": [ - "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_12-/unet/down_blocks.2/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10" - ], - "const_args": [ - "NhwcConv_12_weight_NHWC" - ], - "out_args": [ - "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13", - "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_11", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23" - ], - "const_args": [ - "GroupNorm_11_wts_6_0_9" - ], - "out_args": [ - "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_13-/unet/down_blocks.2/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11" - ], - "const_args": [ - "NhwcConv_13_weight_NHWC" - ], - "out_args": [ - "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_14-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" - ], - "const_args": [ - "NhwcConv_14_weight_NHWC" - ], - "out_args": [ - "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1", - "1", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12", - "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_12", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" - ], - "const_args": [ - "GroupNorm_12_wts_6_1_2" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12" - ], - "const_args": [ - "onnx::MatMul_16850_11_0_20" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" - ], - "const_args": [ - "Attention_4_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" - ], - "const_args": [ - "Attention_4_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" - ], - "const_args": [ - "Attention_4_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_4", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4" - ], - "const_args": [ - "onnx::MatMul_16861_11_0_21" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271", - "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13" - ], - "const_args": [ - "onnx::MatMul_16862" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_4", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272" - ], - "const_args": [ - "onnx::MatMul_16872_11_0_22" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_4_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" - ], - "const_args": [ - "onnx::MatMul_16873_11_0_23_12_4_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_4_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" - ], - "const_args": [ - "onnx::MatMul_16873_11_0_23_12_4_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4" - ], - "const_args": [ - "onnx::MatMul_16874_11_0_24" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" - ], - "const_args": [ - "Attention_5_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" - ], - "const_args": [ - "Attention_5_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" - ], - "const_args": [ - "Attention_5_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_5", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5" - ], - "const_args": [ - "onnx::MatMul_16885_11_0_25" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16" - ], - "const_args": [ - "onnx::MatMul_16886" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_5", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280" - ], - "const_args": [ - "onnx::MatMul_16896_11_0_26" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_5_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" - ], - "const_args": [ - "onnx::MatMul_16897_11_0_27_12_5_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_5_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" - ], - "const_args": [ - "onnx::MatMul_16897_11_0_27_12_5_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5" - ], - "const_args": [ - "onnx::MatMul_16898_11_0_28" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" - ], - "const_args": [ - "Attention_6_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" - ], - "const_args": [ - "Attention_6_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" - ], - "const_args": [ - "Attention_6_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_6", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6" - ], - "const_args": [ - "onnx::MatMul_16909_11_0_29" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19" - ], - "const_args": [ - "onnx::MatMul_16910" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_6", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288" - ], - "const_args": [ - "onnx::MatMul_16920_11_0_30" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_6_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" - ], - "const_args": [ - "onnx::MatMul_16921_11_0_31_12_6_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_6_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" - ], - "const_args": [ - "onnx::MatMul_16921_11_0_31_12_6_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6" - ], - "const_args": [ - "onnx::MatMul_16922_11_0_32" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" - ], - "const_args": [ - "Attention_7_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" - ], - "const_args": [ - "Attention_7_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" - ], - "const_args": [ - "Attention_7_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_7", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7" - ], - "const_args": [ - "onnx::MatMul_16933_11_0_33" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22" - ], - "const_args": [ - "onnx::MatMul_16934" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_7", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296" - ], - "const_args": [ - "onnx::MatMul_16944_11_0_34" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_7_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" - ], - "const_args": [ - "onnx::MatMul_16945_11_0_35_12_7_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_7_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" - ], - "const_args": [ - "onnx::MatMul_16945_11_0_35_12_7_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7" - ], - "const_args": [ - "onnx::MatMul_16946_11_0_36" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" - ], - "const_args": [ - "Attention_8_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" - ], - "const_args": [ - "Attention_8_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" - ], - "const_args": [ - "Attention_8_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_8", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8" - ], - "const_args": [ - "onnx::MatMul_16957_11_0_37" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25" - ], - "const_args": [ - "onnx::MatMul_16958" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_8", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304" - ], - "const_args": [ - "onnx::MatMul_16968_11_0_38" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_8_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" - ], - "const_args": [ - "onnx::MatMul_16969_11_0_39_12_8_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_8_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" - ], - "const_args": [ - "onnx::MatMul_16969_11_0_39_12_8_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8" - ], - "const_args": [ - "onnx::MatMul_16970_11_0_40" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" - ], - "const_args": [ - "Attention_9_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" - ], - "const_args": [ - "Attention_9_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" - ], - "const_args": [ - "Attention_9_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_9", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9" - ], - "const_args": [ - "onnx::MatMul_16981_11_0_41" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28" - ], - "const_args": [ - "onnx::MatMul_16982" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_9", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312" - ], - "const_args": [ - "onnx::MatMul_16992_11_0_42" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_9_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" - ], - "const_args": [ - "onnx::MatMul_16993_11_0_43_12_9_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_9_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" - ], - "const_args": [ - "onnx::MatMul_16993_11_0_43_12_9_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9" - ], - "const_args": [ - "onnx::MatMul_16994_11_0_44" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" - ], - "const_args": [ - "Attention_10_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" - ], - "const_args": [ - "Attention_10_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" - ], - "const_args": [ - "Attention_10_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_10", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10" - ], - "const_args": [ - "onnx::MatMul_17005_11_0_45" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31" - ], - "const_args": [ - "onnx::MatMul_17006" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_10", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320" - ], - "const_args": [ - "onnx::MatMul_17016_11_0_46" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_10_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" - ], - "const_args": [ - "onnx::MatMul_17017_11_0_47_12_10_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_10_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" - ], - "const_args": [ - "onnx::MatMul_17017_11_0_47_12_10_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10" - ], - "const_args": [ - "onnx::MatMul_17018_11_0_48" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" - ], - "const_args": [ - "Attention_11_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" - ], - "const_args": [ - "Attention_11_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" - ], - "const_args": [ - "Attention_11_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_11", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11" - ], - "const_args": [ - "onnx::MatMul_17029_11_0_49" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34" - ], - "const_args": [ - "onnx::MatMul_17030" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_11", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328" - ], - "const_args": [ - "onnx::MatMul_17040_11_0_50" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_11_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" - ], - "const_args": [ - "onnx::MatMul_17041_11_0_51_12_11_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_11_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" - ], - "const_args": [ - "onnx::MatMul_17041_11_0_51_12_11_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11" - ], - "const_args": [ - "onnx::MatMul_17042_11_0_52" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" - ], - "const_args": [ - "Attention_12_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" - ], - "const_args": [ - "Attention_12_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" - ], - "const_args": [ - "Attention_12_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_12", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12" - ], - "const_args": [ - "onnx::MatMul_17053_11_0_53" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37" - ], - "const_args": [ - "onnx::MatMul_17054" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_12", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336" - ], - "const_args": [ - "onnx::MatMul_17064_11_0_54" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_12_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" - ], - "const_args": [ - "onnx::MatMul_17065_11_0_55_12_12_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_12_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" - ], - "const_args": [ - "onnx::MatMul_17065_11_0_55_12_12_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12" - ], - "const_args": [ - "onnx::MatMul_17066_11_0_56" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" - ], - "const_args": [ - "Attention_13_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" - ], - "const_args": [ - "Attention_13_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" - ], - "const_args": [ - "Attention_13_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_13", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13" - ], - "const_args": [ - "onnx::MatMul_17077_11_0_57" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40" - ], - "const_args": [ - "onnx::MatMul_17078" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_13", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344" - ], - "const_args": [ - "onnx::MatMul_17088_11_0_58" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" - ], - "const_args": [ - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_13_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" - ], - "const_args": [ - "onnx::MatMul_17089_11_0_59_12_13_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_13_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" - ], - "const_args": [ - "onnx::MatMul_17089_11_0_59_12_13_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13" - ], - "const_args": [ - "onnx::MatMul_17090_11_0_60" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54" - ], - "const_args": [ - "onnx::MatMul_17091_11_0_61" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348", - "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_13", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" - ], - "const_args": [ - "GroupNorm_13_wts_6_0_10" - ], - "out_args": [ - "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_15-/unet/down_blocks.2/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13" - ], - "const_args": [ - "NhwcConv_15_weight_NHWC" - ], - "out_args": [ - "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15", - "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_14", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56" - ], - "const_args": [ - "GroupNorm_14_wts_6_0_11" - ], - "out_args": [ - "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_16-/unet/down_blocks.2/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14" - ], - "const_args": [ - "NhwcConv_16_weight_NHWC" - ], - "out_args": [ - "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55", - "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_15", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" - ], - "const_args": [ - "GroupNorm_15_wts_6_1_3" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15" - ], - "const_args": [ - "onnx::MatMul_17098_11_0_62" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" - ], - "const_args": [ - "Attention_14_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" - ], - "const_args": [ - "Attention_14_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" - ], - "const_args": [ - "Attention_14_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_14", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14" - ], - "const_args": [ - "onnx::MatMul_17109_11_0_63" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353", - "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43" - ], - "const_args": [ - "onnx::MatMul_17110" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_14", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354" - ], - "const_args": [ - "onnx::MatMul_17120_11_0_64" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_14_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" - ], - "const_args": [ - "onnx::MatMul_17121_11_0_65_12_14_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_14_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" - ], - "const_args": [ - "onnx::MatMul_17121_11_0_65_12_14_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14" - ], - "const_args": [ - "onnx::MatMul_17122_11_0_66" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" - ], - "const_args": [ - "Attention_15_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" - ], - "const_args": [ - "Attention_15_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" - ], - "const_args": [ - "Attention_15_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_15", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15" - ], - "const_args": [ - "onnx::MatMul_17133_11_0_67" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46" - ], - "const_args": [ - "onnx::MatMul_17134" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_15", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362" - ], - "const_args": [ - "onnx::MatMul_17144_11_0_68" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_15_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" - ], - "const_args": [ - "onnx::MatMul_17145_11_0_69_12_15_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_15_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" - ], - "const_args": [ - "onnx::MatMul_17145_11_0_69_12_15_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15" - ], - "const_args": [ - "onnx::MatMul_17146_11_0_70" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" - ], - "const_args": [ - "Attention_16_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" - ], - "const_args": [ - "Attention_16_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" - ], - "const_args": [ - "Attention_16_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_16", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16" - ], - "const_args": [ - "onnx::MatMul_17157_11_0_71" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49" - ], - "const_args": [ - "onnx::MatMul_17158" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_16", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370" - ], - "const_args": [ - "onnx::MatMul_17168_11_0_72" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_16_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" - ], - "const_args": [ - "onnx::MatMul_17169_11_0_73_12_16_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_16_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" - ], - "const_args": [ - "onnx::MatMul_17169_11_0_73_12_16_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16" - ], - "const_args": [ - "onnx::MatMul_17170_11_0_74" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" - ], - "const_args": [ - "Attention_17_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" - ], - "const_args": [ - "Attention_17_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" - ], - "const_args": [ - "Attention_17_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_17", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17" - ], - "const_args": [ - "onnx::MatMul_17181_11_0_75" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52" - ], - "const_args": [ - "onnx::MatMul_17182" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_17", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378" - ], - "const_args": [ - "onnx::MatMul_17192_11_0_76" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_17_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" - ], - "const_args": [ - "onnx::MatMul_17193_11_0_77_12_17_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_17_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" - ], - "const_args": [ - "onnx::MatMul_17193_11_0_77_12_17_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17" - ], - "const_args": [ - "onnx::MatMul_17194_11_0_78" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" - ], - "const_args": [ - "Attention_18_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" - ], - "const_args": [ - "Attention_18_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" - ], - "const_args": [ - "Attention_18_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_18", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18" - ], - "const_args": [ - "onnx::MatMul_17205_11_0_79" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55" - ], - "const_args": [ - "onnx::MatMul_17206" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_18", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386" - ], - "const_args": [ - "onnx::MatMul_17216_11_0_80" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_18_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" - ], - "const_args": [ - "onnx::MatMul_17217_11_0_81_12_18_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_18_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" - ], - "const_args": [ - "onnx::MatMul_17217_11_0_81_12_18_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18" - ], - "const_args": [ - "onnx::MatMul_17218_11_0_82" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" - ], - "const_args": [ - "Attention_19_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" - ], - "const_args": [ - "Attention_19_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" - ], - "const_args": [ - "Attention_19_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_19", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19" - ], - "const_args": [ - "onnx::MatMul_17229_11_0_83" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58" - ], - "const_args": [ - "onnx::MatMul_17230" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_19", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394" - ], - "const_args": [ - "onnx::MatMul_17240_11_0_84" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_19_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" - ], - "const_args": [ - "onnx::MatMul_17241_11_0_85_12_19_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_19_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" - ], - "const_args": [ - "onnx::MatMul_17241_11_0_85_12_19_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19" - ], - "const_args": [ - "onnx::MatMul_17242_11_0_86" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" - ], - "const_args": [ - "Attention_20_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" - ], - "const_args": [ - "Attention_20_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" - ], - "const_args": [ - "Attention_20_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_20", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20" - ], - "const_args": [ - "onnx::MatMul_17253_11_0_87" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61" - ], - "const_args": [ - "onnx::MatMul_17254" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_20", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402" - ], - "const_args": [ - "onnx::MatMul_17264_11_0_88" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_20_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" - ], - "const_args": [ - "onnx::MatMul_17265_11_0_89_12_20_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_20_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" - ], - "const_args": [ - "onnx::MatMul_17265_11_0_89_12_20_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20" - ], - "const_args": [ - "onnx::MatMul_17266_11_0_90" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" - ], - "const_args": [ - "Attention_21_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" - ], - "const_args": [ - "Attention_21_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" - ], - "const_args": [ - "Attention_21_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_21", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21" - ], - "const_args": [ - "onnx::MatMul_17277_11_0_91" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64" - ], - "const_args": [ - "onnx::MatMul_17278" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_21", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410" - ], - "const_args": [ - "onnx::MatMul_17288_11_0_92" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_21_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" - ], - "const_args": [ - "onnx::MatMul_17289_11_0_93_12_21_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_21_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" - ], - "const_args": [ - "onnx::MatMul_17289_11_0_93_12_21_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21" - ], - "const_args": [ - "onnx::MatMul_17290_11_0_94" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" - ], - "const_args": [ - "Attention_22_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" - ], - "const_args": [ - "Attention_22_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" - ], - "const_args": [ - "Attention_22_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_22", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22" - ], - "const_args": [ - "onnx::MatMul_17301_11_0_95" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67" - ], - "const_args": [ - "onnx::MatMul_17302" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_22", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418" - ], - "const_args": [ - "onnx::MatMul_17312_11_0_96" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_22_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" - ], - "const_args": [ - "onnx::MatMul_17313_11_0_97_12_22_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_22_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" - ], - "const_args": [ - "onnx::MatMul_17313_11_0_97_12_22_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22" - ], - "const_args": [ - "onnx::MatMul_17314_11_0_98" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" - ], - "const_args": [ - "Attention_23_qkv_weight_key" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" - ], - "const_args": [ - "Attention_23_qkv_weight_query" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" - ], - "const_args": [ - "Attention_23_qkv_weight_value" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_23", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23" - ], - "const_args": [ - "onnx::MatMul_17325_11_0_99" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70" - ], - "const_args": [ - "onnx::MatMul_17326" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_23", - "type": "SDFlatMHA", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47" - ], - "const_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426" - ], - "const_args": [ - "onnx::MatMul_17336_11_0_100" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" - ], - "const_args": [ - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_23_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" - ], - "const_args": [ - "onnx::MatMul_17337_11_0_101_12_23_0" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_23_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" - ], - "const_args": [ - "onnx::MatMul_17337_11_0_101_12_23_1" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23" - ], - "const_args": [ - "onnx::MatMul_17338_11_0_102" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87" - ], - "const_args": [ - "onnx::MatMul_17339_11_0_103" - ], - "out_args": [ - "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/down_blocks.2/attentions.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430", - "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" - ], - "const_args": [], - "out_args": [ - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_16", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" - ], - "const_args": [ - "GroupNorm_16_wts_6_0_12" - ], - "out_args": [ - "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_17-/unet/mid_block/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16" - ], - "const_args": [ - "NhwcConv_17_weight_NHWC" - ], - "out_args": [ - "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17", - "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_17", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89" - ], - "const_args": [ - "GroupNorm_17_wts_6_0_13" - ], - "out_args": [ - "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_18-/unet/mid_block/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17" - ], - "const_args": [ - "NhwcConv_18_weight_NHWC" - ], - "out_args": [ - "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88", - "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_18", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" - ], - "const_args": [ - "GroupNorm_18_wts_6_1_4" - ], - "out_args": [ - "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18" - ], - "const_args": [ - "onnx::MatMul_17346_11_0_104" - ], - "out_args": [ - "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" - ], - "const_args": [ - "Attention_24_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" - ], - "const_args": [ - "Attention_24_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" - ], - "const_args": [ - "Attention_24_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_24", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24" - ], - "const_args": [ - "onnx::MatMul_17357_11_0_105" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435", - "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73" - ], - "const_args": [ - "onnx::MatMul_17358" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_24", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436" - ], - "const_args": [ - "onnx::MatMul_17368_11_0_106" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_24_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" - ], - "const_args": [ - "onnx::MatMul_17369_11_0_107_12_24_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_24_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" - ], - "const_args": [ - "onnx::MatMul_17369_11_0_107_12_24_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438", - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24" - ], - "const_args": [ - "onnx::MatMul_17370_11_0_108" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" - ], - "const_args": [ - "Attention_25_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" - ], - "const_args": [ - "Attention_25_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" - ], - "const_args": [ - "Attention_25_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_25", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25" - ], - "const_args": [ - "onnx::MatMul_17381_11_0_109" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76" - ], - "const_args": [ - "onnx::MatMul_17382" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_25", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444" - ], - "const_args": [ - "onnx::MatMul_17392_11_0_110" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_25_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" - ], - "const_args": [ - "onnx::MatMul_17393_11_0_111_12_25_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_25_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" - ], - "const_args": [ - "onnx::MatMul_17393_11_0_111_12_25_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446", - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25" - ], - "const_args": [ - "onnx::MatMul_17394_11_0_112" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" - ], - "const_args": [ - "Attention_26_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" - ], - "const_args": [ - "Attention_26_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" - ], - "const_args": [ - "Attention_26_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_26", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26" - ], - "const_args": [ - "onnx::MatMul_17405_11_0_113" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79" - ], - "const_args": [ - "onnx::MatMul_17406" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_26", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452" - ], - "const_args": [ - "onnx::MatMul_17416_11_0_114" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_26_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" - ], - "const_args": [ - "onnx::MatMul_17417_11_0_115_12_26_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_26_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" - ], - "const_args": [ - "onnx::MatMul_17417_11_0_115_12_26_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454", - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26" - ], - "const_args": [ - "onnx::MatMul_17418_11_0_116" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" - ], - "const_args": [ - "Attention_27_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" - ], - "const_args": [ - "Attention_27_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" - ], - "const_args": [ - "Attention_27_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_27", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27" - ], - "const_args": [ - "onnx::MatMul_17429_11_0_117" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82" - ], - "const_args": [ - "onnx::MatMul_17430" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_27", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460" - ], - "const_args": [ - "onnx::MatMul_17440_11_0_118" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_27_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" - ], - "const_args": [ - "onnx::MatMul_17441_11_0_119_12_27_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_27_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" - ], - "const_args": [ - "onnx::MatMul_17441_11_0_119_12_27_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462", - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27" - ], - "const_args": [ - "onnx::MatMul_17442_11_0_120" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" - ], - "const_args": [ - "Attention_28_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" - ], - "const_args": [ - "Attention_28_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" - ], - "const_args": [ - "Attention_28_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_28", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28" - ], - "const_args": [ - "onnx::MatMul_17453_11_0_121" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85" - ], - "const_args": [ - "onnx::MatMul_17454" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_28", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468" - ], - "const_args": [ - "onnx::MatMul_17464_11_0_122" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_28_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" - ], - "const_args": [ - "onnx::MatMul_17465_11_0_123_12_28_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_28_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" - ], - "const_args": [ - "onnx::MatMul_17465_11_0_123_12_28_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470", - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28" - ], - "const_args": [ - "onnx::MatMul_17466_11_0_124" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" - ], - "const_args": [ - "Attention_29_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" - ], - "const_args": [ - "Attention_29_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" - ], - "const_args": [ - "Attention_29_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_29", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29" - ], - "const_args": [ - "onnx::MatMul_17477_11_0_125" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88" - ], - "const_args": [ - "onnx::MatMul_17478" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_29", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476" - ], - "const_args": [ - "onnx::MatMul_17488_11_0_126" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_29_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" - ], - "const_args": [ - "onnx::MatMul_17489_11_0_127_12_29_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_29_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" - ], - "const_args": [ - "onnx::MatMul_17489_11_0_127_12_29_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478", - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29" - ], - "const_args": [ - "onnx::MatMul_17490_11_0_128" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" - ], - "const_args": [ - "Attention_30_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" - ], - "const_args": [ - "Attention_30_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" - ], - "const_args": [ - "Attention_30_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_30", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30" - ], - "const_args": [ - "onnx::MatMul_17501_11_0_129" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91" - ], - "const_args": [ - "onnx::MatMul_17502" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_30", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484" - ], - "const_args": [ - "onnx::MatMul_17512_11_0_130" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_30_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" - ], - "const_args": [ - "onnx::MatMul_17513_11_0_131_12_30_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_30_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" - ], - "const_args": [ - "onnx::MatMul_17513_11_0_131_12_30_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486", - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30" - ], - "const_args": [ - "onnx::MatMul_17514_11_0_132" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" - ], - "const_args": [ - "Attention_31_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" - ], - "const_args": [ - "Attention_31_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" - ], - "const_args": [ - "Attention_31_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_31", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31" - ], - "const_args": [ - "onnx::MatMul_17525_11_0_133" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94" - ], - "const_args": [ - "onnx::MatMul_17526" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_31", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492" - ], - "const_args": [ - "onnx::MatMul_17536_11_0_134" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_31_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" - ], - "const_args": [ - "onnx::MatMul_17537_11_0_135_12_31_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_31_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" - ], - "const_args": [ - "onnx::MatMul_17537_11_0_135_12_31_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494", - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31" - ], - "const_args": [ - "onnx::MatMul_17538_11_0_136" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" - ], - "const_args": [ - "Attention_32_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" - ], - "const_args": [ - "Attention_32_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" - ], - "const_args": [ - "Attention_32_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_32", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32" - ], - "const_args": [ - "onnx::MatMul_17549_11_0_137" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97" - ], - "const_args": [ - "onnx::MatMul_17550" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_32", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500" - ], - "const_args": [ - "onnx::MatMul_17560_11_0_138" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_32_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" - ], - "const_args": [ - "onnx::MatMul_17561_11_0_139_12_32_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_32_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" - ], - "const_args": [ - "onnx::MatMul_17561_11_0_139_12_32_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502", - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32" - ], - "const_args": [ - "onnx::MatMul_17562_11_0_140" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" - ], - "const_args": [ - "Attention_33_qkv_weight_key" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" - ], - "const_args": [ - "Attention_33_qkv_weight_query" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" - ], - "const_args": [ - "Attention_33_qkv_weight_value" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_33", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33" - ], - "const_args": [ - "onnx::MatMul_17573_11_0_141" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100" - ], - "const_args": [ - "onnx::MatMul_17574" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_33", - "type": "SDFlatMHA", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67" - ], - "const_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508" - ], - "const_args": [ - "onnx::MatMul_17584_11_0_142" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508", - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" - ], - "const_args": [ - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_33_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" - ], - "const_args": [ - "onnx::MatMul_17585_11_0_143_12_33_0" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_33_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" - ], - "const_args": [ - "onnx::MatMul_17585_11_0_143_12_33_1" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510", - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33" - ], - "const_args": [ - "onnx::MatMul_17586_11_0_144" - ], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511", - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120" - ], - "const_args": [ - "onnx::MatMul_17587_11_0_145" - ], - "out_args": [ - "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/attentions.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512", - "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_19", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121" - ], - "const_args": [ - "GroupNorm_19_wts_6_0_14" - ], - "out_args": [ - "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_19-/unet/mid_block/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19" - ], - "const_args": [ - "NhwcConv_19_weight_NHWC" - ], - "out_args": [ - "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19", - "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_20", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122" - ], - "const_args": [ - "GroupNorm_20_wts_6_0_15" - ], - "out_args": [ - "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_20-/unet/mid_block/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20" - ], - "const_args": [ - "NhwcConv_20_weight_NHWC" - ], - "out_args": [ - "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/mid_block/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121", - "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20" - ], - "const_args": [], - "out_args": [ - "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/Concat", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123", - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_21", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" - ], - "const_args": [ - "GroupNorm_21_wts_6_0_16" - ], - "out_args": [ - "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "5120" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_21-/unet/up_blocks.0/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21" - ], - "const_args": [ - "NhwcConv_21_weight_NHWC" - ], - "out_args": [ - "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "2560" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22", - "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_22", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124" - ], - "const_args": [ - "GroupNorm_22_wts_6_0_17" - ], - "out_args": [ - "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_22-/unet/up_blocks.0/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22" - ], - "const_args": [ - "NhwcConv_22_weight_NHWC" - ], - "out_args": [ - "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_23-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" - ], - "const_args": [ - "NhwcConv_23_weight_NHWC" - ], - "out_args": [ - "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1", - "1", - "2560" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21", - "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_23", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" - ], - "const_args": [ - "GroupNorm_23_wts_6_1_5" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23" - ], - "const_args": [ - "onnx::MatMul_17598_11_0_146" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" - ], - "const_args": [ - "Attention_34_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" - ], - "const_args": [ - "Attention_34_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" - ], - "const_args": [ - "Attention_34_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_34", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34" - ], - "const_args": [ - "onnx::MatMul_17609_11_0_147" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517", - "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103" - ], - "const_args": [ - "onnx::MatMul_17610" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_34", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518" - ], - "const_args": [ - "onnx::MatMul_17620_11_0_148" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_34_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" - ], - "const_args": [ - "onnx::MatMul_17621_11_0_149_12_34_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_34_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" - ], - "const_args": [ - "onnx::MatMul_17621_11_0_149_12_34_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34" - ], - "const_args": [ - "onnx::MatMul_17622_11_0_150" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" - ], - "const_args": [ - "Attention_35_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" - ], - "const_args": [ - "Attention_35_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" - ], - "const_args": [ - "Attention_35_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_35", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35" - ], - "const_args": [ - "onnx::MatMul_17633_11_0_151" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106" - ], - "const_args": [ - "onnx::MatMul_17634" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_35", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526" - ], - "const_args": [ - "onnx::MatMul_17644_11_0_152" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_35_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" - ], - "const_args": [ - "onnx::MatMul_17645_11_0_153_12_35_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_35_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" - ], - "const_args": [ - "onnx::MatMul_17645_11_0_153_12_35_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35" - ], - "const_args": [ - "onnx::MatMul_17646_11_0_154" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" - ], - "const_args": [ - "Attention_36_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" - ], - "const_args": [ - "Attention_36_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" - ], - "const_args": [ - "Attention_36_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_36", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36" - ], - "const_args": [ - "onnx::MatMul_17657_11_0_155" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109" - ], - "const_args": [ - "onnx::MatMul_17658" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_36", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534" - ], - "const_args": [ - "onnx::MatMul_17668_11_0_156" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_36_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" - ], - "const_args": [ - "onnx::MatMul_17669_11_0_157_12_36_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_36_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" - ], - "const_args": [ - "onnx::MatMul_17669_11_0_157_12_36_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36" - ], - "const_args": [ - "onnx::MatMul_17670_11_0_158" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" - ], - "const_args": [ - "Attention_37_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" - ], - "const_args": [ - "Attention_37_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" - ], - "const_args": [ - "Attention_37_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_37", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37" - ], - "const_args": [ - "onnx::MatMul_17681_11_0_159" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112" - ], - "const_args": [ - "onnx::MatMul_17682" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_37", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542" - ], - "const_args": [ - "onnx::MatMul_17692_11_0_160" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_37_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" - ], - "const_args": [ - "onnx::MatMul_17693_11_0_161_12_37_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_37_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" - ], - "const_args": [ - "onnx::MatMul_17693_11_0_161_12_37_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37" - ], - "const_args": [ - "onnx::MatMul_17694_11_0_162" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" - ], - "const_args": [ - "Attention_38_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" - ], - "const_args": [ - "Attention_38_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" - ], - "const_args": [ - "Attention_38_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_38", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38" - ], - "const_args": [ - "onnx::MatMul_17705_11_0_163" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115" - ], - "const_args": [ - "onnx::MatMul_17706" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_38", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550" - ], - "const_args": [ - "onnx::MatMul_17716_11_0_164" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_38_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" - ], - "const_args": [ - "onnx::MatMul_17717_11_0_165_12_38_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_38_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" - ], - "const_args": [ - "onnx::MatMul_17717_11_0_165_12_38_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38" - ], - "const_args": [ - "onnx::MatMul_17718_11_0_166" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" - ], - "const_args": [ - "Attention_39_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" - ], - "const_args": [ - "Attention_39_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" - ], - "const_args": [ - "Attention_39_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_39", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39" - ], - "const_args": [ - "onnx::MatMul_17729_11_0_167" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118" - ], - "const_args": [ - "onnx::MatMul_17730" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_39", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558" - ], - "const_args": [ - "onnx::MatMul_17740_11_0_168" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_39_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" - ], - "const_args": [ - "onnx::MatMul_17741_11_0_169_12_39_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_39_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" - ], - "const_args": [ - "onnx::MatMul_17741_11_0_169_12_39_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39" - ], - "const_args": [ - "onnx::MatMul_17742_11_0_170" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" - ], - "const_args": [ - "Attention_40_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" - ], - "const_args": [ - "Attention_40_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" - ], - "const_args": [ - "Attention_40_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_40", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40" - ], - "const_args": [ - "onnx::MatMul_17753_11_0_171" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121" - ], - "const_args": [ - "onnx::MatMul_17754" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_40", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566" - ], - "const_args": [ - "onnx::MatMul_17764_11_0_172" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_40_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" - ], - "const_args": [ - "onnx::MatMul_17765_11_0_173_12_40_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_40_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" - ], - "const_args": [ - "onnx::MatMul_17765_11_0_173_12_40_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40" - ], - "const_args": [ - "onnx::MatMul_17766_11_0_174" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" - ], - "const_args": [ - "Attention_41_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" - ], - "const_args": [ - "Attention_41_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" - ], - "const_args": [ - "Attention_41_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_41", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41" - ], - "const_args": [ - "onnx::MatMul_17777_11_0_175" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124" - ], - "const_args": [ - "onnx::MatMul_17778" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_41", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574" - ], - "const_args": [ - "onnx::MatMul_17788_11_0_176" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_41_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" - ], - "const_args": [ - "onnx::MatMul_17789_11_0_177_12_41_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_41_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" - ], - "const_args": [ - "onnx::MatMul_17789_11_0_177_12_41_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41" - ], - "const_args": [ - "onnx::MatMul_17790_11_0_178" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" - ], - "const_args": [ - "Attention_42_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" - ], - "const_args": [ - "Attention_42_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" - ], - "const_args": [ - "Attention_42_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_42", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42" - ], - "const_args": [ - "onnx::MatMul_17801_11_0_179" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127" - ], - "const_args": [ - "onnx::MatMul_17802" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_42", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582" - ], - "const_args": [ - "onnx::MatMul_17812_11_0_180" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_42_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" - ], - "const_args": [ - "onnx::MatMul_17813_11_0_181_12_42_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_42_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" - ], - "const_args": [ - "onnx::MatMul_17813_11_0_181_12_42_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42" - ], - "const_args": [ - "onnx::MatMul_17814_11_0_182" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" - ], - "const_args": [ - "Attention_43_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" - ], - "const_args": [ - "Attention_43_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" - ], - "const_args": [ - "Attention_43_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_43", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43" - ], - "const_args": [ - "onnx::MatMul_17825_11_0_183" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130" - ], - "const_args": [ - "onnx::MatMul_17826" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_43", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590" - ], - "const_args": [ - "onnx::MatMul_17836_11_0_184" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" - ], - "const_args": [ - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_43_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" - ], - "const_args": [ - "onnx::MatMul_17837_11_0_185_12_43_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_43_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" - ], - "const_args": [ - "onnx::MatMul_17837_11_0_185_12_43_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43" - ], - "const_args": [ - "onnx::MatMul_17838_11_0_186" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155" - ], - "const_args": [ - "onnx::MatMul_17839_11_0_187" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594", - "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/Concat_1", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156", - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_24", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" - ], - "const_args": [ - "GroupNorm_24_wts_6_0_18" - ], - "out_args": [ - "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "5120" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_24-/unet/up_blocks.0/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24" - ], - "const_args": [ - "NhwcConv_24_weight_NHWC" - ], - "out_args": [ - "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "2560" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25", - "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_25", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157" - ], - "const_args": [ - "GroupNorm_25_wts_6_0_19" - ], - "out_args": [ - "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_25-/unet/up_blocks.0/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25" - ], - "const_args": [ - "NhwcConv_25_weight_NHWC" - ], - "out_args": [ - "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_26-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" - ], - "const_args": [ - "NhwcConv_26_weight_NHWC" - ], - "out_args": [ - "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1", - "1", - "2560" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24", - "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_26", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" - ], - "const_args": [ - "GroupNorm_26_wts_6_1_6" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26" - ], - "const_args": [ - "onnx::MatMul_17846_11_0_188" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" - ], - "const_args": [ - "Attention_44_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" - ], - "const_args": [ - "Attention_44_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" - ], - "const_args": [ - "Attention_44_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_44", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44" - ], - "const_args": [ - "onnx::MatMul_17857_11_0_189" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599", - "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133" - ], - "const_args": [ - "onnx::MatMul_17858" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_44", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600" - ], - "const_args": [ - "onnx::MatMul_17868_11_0_190" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_44_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" - ], - "const_args": [ - "onnx::MatMul_17869_11_0_191_12_44_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_44_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" - ], - "const_args": [ - "onnx::MatMul_17869_11_0_191_12_44_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44" - ], - "const_args": [ - "onnx::MatMul_17870_11_0_192" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" - ], - "const_args": [ - "Attention_45_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" - ], - "const_args": [ - "Attention_45_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" - ], - "const_args": [ - "Attention_45_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_45", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45" - ], - "const_args": [ - "onnx::MatMul_17881_11_0_193" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136" - ], - "const_args": [ - "onnx::MatMul_17882" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_45", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608" - ], - "const_args": [ - "onnx::MatMul_17892_11_0_194" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_45_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" - ], - "const_args": [ - "onnx::MatMul_17893_11_0_195_12_45_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_45_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" - ], - "const_args": [ - "onnx::MatMul_17893_11_0_195_12_45_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45" - ], - "const_args": [ - "onnx::MatMul_17894_11_0_196" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" - ], - "const_args": [ - "Attention_46_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" - ], - "const_args": [ - "Attention_46_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" - ], - "const_args": [ - "Attention_46_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_46", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46" - ], - "const_args": [ - "onnx::MatMul_17905_11_0_197" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139" - ], - "const_args": [ - "onnx::MatMul_17906" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_46", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616" - ], - "const_args": [ - "onnx::MatMul_17916_11_0_198" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_46_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" - ], - "const_args": [ - "onnx::MatMul_17917_11_0_199_12_46_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_46_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" - ], - "const_args": [ - "onnx::MatMul_17917_11_0_199_12_46_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46" - ], - "const_args": [ - "onnx::MatMul_17918_11_0_200" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" - ], - "const_args": [ - "Attention_47_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" - ], - "const_args": [ - "Attention_47_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" - ], - "const_args": [ - "Attention_47_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_47", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47" - ], - "const_args": [ - "onnx::MatMul_17929_11_0_201" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142" - ], - "const_args": [ - "onnx::MatMul_17930" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_47", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624" - ], - "const_args": [ - "onnx::MatMul_17940_11_0_202" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_47_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" - ], - "const_args": [ - "onnx::MatMul_17941_11_0_203_12_47_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_47_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" - ], - "const_args": [ - "onnx::MatMul_17941_11_0_203_12_47_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47" - ], - "const_args": [ - "onnx::MatMul_17942_11_0_204" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" - ], - "const_args": [ - "Attention_48_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" - ], - "const_args": [ - "Attention_48_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" - ], - "const_args": [ - "Attention_48_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_48", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48" - ], - "const_args": [ - "onnx::MatMul_17953_11_0_205" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145" - ], - "const_args": [ - "onnx::MatMul_17954" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_48", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632" - ], - "const_args": [ - "onnx::MatMul_17964_11_0_206" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_48_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" - ], - "const_args": [ - "onnx::MatMul_17965_11_0_207_12_48_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_48_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" - ], - "const_args": [ - "onnx::MatMul_17965_11_0_207_12_48_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48" - ], - "const_args": [ - "onnx::MatMul_17966_11_0_208" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" - ], - "const_args": [ - "Attention_49_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" - ], - "const_args": [ - "Attention_49_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" - ], - "const_args": [ - "Attention_49_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_49", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49" - ], - "const_args": [ - "onnx::MatMul_17977_11_0_209" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148" - ], - "const_args": [ - "onnx::MatMul_17978" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_49", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640" - ], - "const_args": [ - "onnx::MatMul_17988_11_0_210" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_49_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" - ], - "const_args": [ - "onnx::MatMul_17989_11_0_211_12_49_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_49_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" - ], - "const_args": [ - "onnx::MatMul_17989_11_0_211_12_49_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49" - ], - "const_args": [ - "onnx::MatMul_17990_11_0_212" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" - ], - "const_args": [ - "Attention_50_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" - ], - "const_args": [ - "Attention_50_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" - ], - "const_args": [ - "Attention_50_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_50", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50" - ], - "const_args": [ - "onnx::MatMul_18001_11_0_213" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151" - ], - "const_args": [ - "onnx::MatMul_18002" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_50", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648" - ], - "const_args": [ - "onnx::MatMul_18012_11_0_214" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_50_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" - ], - "const_args": [ - "onnx::MatMul_18013_11_0_215_12_50_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_50_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" - ], - "const_args": [ - "onnx::MatMul_18013_11_0_215_12_50_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50" - ], - "const_args": [ - "onnx::MatMul_18014_11_0_216" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" - ], - "const_args": [ - "Attention_51_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" - ], - "const_args": [ - "Attention_51_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" - ], - "const_args": [ - "Attention_51_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_51", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51" - ], - "const_args": [ - "onnx::MatMul_18025_11_0_217" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154" - ], - "const_args": [ - "onnx::MatMul_18026" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_51", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656" - ], - "const_args": [ - "onnx::MatMul_18036_11_0_218" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_51_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" - ], - "const_args": [ - "onnx::MatMul_18037_11_0_219_12_51_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_51_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" - ], - "const_args": [ - "onnx::MatMul_18037_11_0_219_12_51_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51" - ], - "const_args": [ - "onnx::MatMul_18038_11_0_220" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" - ], - "const_args": [ - "Attention_52_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" - ], - "const_args": [ - "Attention_52_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" - ], - "const_args": [ - "Attention_52_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_52", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52" - ], - "const_args": [ - "onnx::MatMul_18049_11_0_221" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157" - ], - "const_args": [ - "onnx::MatMul_18050" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_52", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664" - ], - "const_args": [ - "onnx::MatMul_18060_11_0_222" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_52_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" - ], - "const_args": [ - "onnx::MatMul_18061_11_0_223_12_52_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_52_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" - ], - "const_args": [ - "onnx::MatMul_18061_11_0_223_12_52_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52" - ], - "const_args": [ - "onnx::MatMul_18062_11_0_224" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" - ], - "const_args": [ - "Attention_53_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" - ], - "const_args": [ - "Attention_53_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" - ], - "const_args": [ - "Attention_53_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_53", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53" - ], - "const_args": [ - "onnx::MatMul_18073_11_0_225" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160" - ], - "const_args": [ - "onnx::MatMul_18074" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_53", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672" - ], - "const_args": [ - "onnx::MatMul_18084_11_0_226" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" - ], - "const_args": [ - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_53_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" - ], - "const_args": [ - "onnx::MatMul_18085_11_0_227_12_53_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_53_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" - ], - "const_args": [ - "onnx::MatMul_18085_11_0_227_12_53_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53" - ], - "const_args": [ - "onnx::MatMul_18086_11_0_228" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188" - ], - "const_args": [ - "onnx::MatMul_18087_11_0_229" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676", - "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/Concat_2", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189", - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1920" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_27", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" - ], - "const_args": [ - "GroupNorm_27_wts_6_0_20" - ], - "out_args": [ - "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1920" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "3840" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_27-/unet/up_blocks.0/resnets.2/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27" - ], - "const_args": [ - "NhwcConv_27_weight_NHWC" - ], - "out_args": [ - "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1920" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28", - "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_28", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190" - ], - "const_args": [ - "GroupNorm_28_wts_6_0_21" - ], - "out_args": [ - "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_28-/unet/up_blocks.0/resnets.2/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28" - ], - "const_args": [ - "NhwcConv_28_weight_NHWC" - ], - "out_args": [ - "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_29-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" - ], - "const_args": [ - "NhwcConv_29_weight_NHWC" - ], - "out_args": [ - "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1", - "1", - "1920" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/resnets.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27", - "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_29", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" - ], - "const_args": [ - "GroupNorm_29_wts_6_1_7" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29" - ], - "const_args": [ - "onnx::MatMul_18094_11_0_230" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" - ], - "const_args": [ - "Attention_54_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" - ], - "const_args": [ - "Attention_54_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" - ], - "const_args": [ - "Attention_54_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_54", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54" - ], - "const_args": [ - "onnx::MatMul_18105_11_0_231" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681", - "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163" - ], - "const_args": [ - "onnx::MatMul_18106" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_54", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682" - ], - "const_args": [ - "onnx::MatMul_18116_11_0_232" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_54_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" - ], - "const_args": [ - "onnx::MatMul_18117_11_0_233_12_54_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_54_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" - ], - "const_args": [ - "onnx::MatMul_18117_11_0_233_12_54_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54" - ], - "const_args": [ - "onnx::MatMul_18118_11_0_234" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" - ], - "const_args": [ - "Attention_55_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" - ], - "const_args": [ - "Attention_55_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" - ], - "const_args": [ - "Attention_55_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_55", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55" - ], - "const_args": [ - "onnx::MatMul_18129_11_0_235" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166" - ], - "const_args": [ - "onnx::MatMul_18130" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_55", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690" - ], - "const_args": [ - "onnx::MatMul_18140_11_0_236" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_55_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" - ], - "const_args": [ - "onnx::MatMul_18141_11_0_237_12_55_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_55_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" - ], - "const_args": [ - "onnx::MatMul_18141_11_0_237_12_55_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55" - ], - "const_args": [ - "onnx::MatMul_18142_11_0_238" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" - ], - "const_args": [ - "Attention_56_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" - ], - "const_args": [ - "Attention_56_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" - ], - "const_args": [ - "Attention_56_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_56", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56" - ], - "const_args": [ - "onnx::MatMul_18153_11_0_239" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169" - ], - "const_args": [ - "onnx::MatMul_18154" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_56", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698" - ], - "const_args": [ - "onnx::MatMul_18164_11_0_240" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_56_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" - ], - "const_args": [ - "onnx::MatMul_18165_11_0_241_12_56_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_56_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" - ], - "const_args": [ - "onnx::MatMul_18165_11_0_241_12_56_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56" - ], - "const_args": [ - "onnx::MatMul_18166_11_0_242" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" - ], - "const_args": [ - "Attention_57_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" - ], - "const_args": [ - "Attention_57_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" - ], - "const_args": [ - "Attention_57_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_57", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57" - ], - "const_args": [ - "onnx::MatMul_18177_11_0_243" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172" - ], - "const_args": [ - "onnx::MatMul_18178" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_57", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706" - ], - "const_args": [ - "onnx::MatMul_18188_11_0_244" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_57_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" - ], - "const_args": [ - "onnx::MatMul_18189_11_0_245_12_57_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_57_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" - ], - "const_args": [ - "onnx::MatMul_18189_11_0_245_12_57_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57" - ], - "const_args": [ - "onnx::MatMul_18190_11_0_246" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" - ], - "const_args": [ - "Attention_58_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" - ], - "const_args": [ - "Attention_58_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" - ], - "const_args": [ - "Attention_58_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_58", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58" - ], - "const_args": [ - "onnx::MatMul_18201_11_0_247" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175" - ], - "const_args": [ - "onnx::MatMul_18202" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_58", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714" - ], - "const_args": [ - "onnx::MatMul_18212_11_0_248" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_58_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" - ], - "const_args": [ - "onnx::MatMul_18213_11_0_249_12_58_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_58_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" - ], - "const_args": [ - "onnx::MatMul_18213_11_0_249_12_58_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58" - ], - "const_args": [ - "onnx::MatMul_18214_11_0_250" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" - ], - "const_args": [ - "Attention_59_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" - ], - "const_args": [ - "Attention_59_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" - ], - "const_args": [ - "Attention_59_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_59", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59" - ], - "const_args": [ - "onnx::MatMul_18225_11_0_251" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178" - ], - "const_args": [ - "onnx::MatMul_18226" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_59", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722" - ], - "const_args": [ - "onnx::MatMul_18236_11_0_252" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_59_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" - ], - "const_args": [ - "onnx::MatMul_18237_11_0_253_12_59_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_59_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" - ], - "const_args": [ - "onnx::MatMul_18237_11_0_253_12_59_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59" - ], - "const_args": [ - "onnx::MatMul_18238_11_0_254" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" - ], - "const_args": [ - "Attention_60_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" - ], - "const_args": [ - "Attention_60_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" - ], - "const_args": [ - "Attention_60_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_60", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60" - ], - "const_args": [ - "onnx::MatMul_18249_11_0_255" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181" - ], - "const_args": [ - "onnx::MatMul_18250" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_60", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730" - ], - "const_args": [ - "onnx::MatMul_18260_11_0_256" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_60_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" - ], - "const_args": [ - "onnx::MatMul_18261_11_0_257_12_60_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_60_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" - ], - "const_args": [ - "onnx::MatMul_18261_11_0_257_12_60_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60" - ], - "const_args": [ - "onnx::MatMul_18262_11_0_258" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" - ], - "const_args": [ - "Attention_61_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" - ], - "const_args": [ - "Attention_61_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" - ], - "const_args": [ - "Attention_61_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_61", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61" - ], - "const_args": [ - "onnx::MatMul_18273_11_0_259" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184" - ], - "const_args": [ - "onnx::MatMul_18274" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_61", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738" - ], - "const_args": [ - "onnx::MatMul_18284_11_0_260" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_61_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" - ], - "const_args": [ - "onnx::MatMul_18285_11_0_261_12_61_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_61_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" - ], - "const_args": [ - "onnx::MatMul_18285_11_0_261_12_61_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61" - ], - "const_args": [ - "onnx::MatMul_18286_11_0_262" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" - ], - "const_args": [ - "Attention_62_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" - ], - "const_args": [ - "Attention_62_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" - ], - "const_args": [ - "Attention_62_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_62", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62" - ], - "const_args": [ - "onnx::MatMul_18297_11_0_263" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187" - ], - "const_args": [ - "onnx::MatMul_18298" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_62", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746" - ], - "const_args": [ - "onnx::MatMul_18308_11_0_264" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_62_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" - ], - "const_args": [ - "onnx::MatMul_18309_11_0_265_12_62_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_62_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" - ], - "const_args": [ - "onnx::MatMul_18309_11_0_265_12_62_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62" - ], - "const_args": [ - "onnx::MatMul_18310_11_0_266" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" - ], - "const_args": [ - "Attention_63_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" - ], - "const_args": [ - "Attention_63_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" - ], - "const_args": [ - "Attention_63_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_63", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "1024" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63" - ], - "const_args": [ - "onnx::MatMul_18321_11_0_267" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190" - ], - "const_args": [ - "onnx::MatMul_18322" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "20" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "20", - "1024", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_63", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "20" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127" - ], - "const_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754" - ], - "const_args": [ - "onnx::MatMul_18332_11_0_268" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" - ], - "const_args": [ - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_63_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" - ], - "const_args": [ - "onnx::MatMul_18333_11_0_269_12_63_0" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_63_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" - ], - "const_args": [ - "onnx::MatMul_18333_11_0_269_12_63_1" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "5120" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63" - ], - "const_args": [ - "onnx::MatMul_18334_11_0_270" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "5120" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "5120", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221" - ], - "const_args": [ - "onnx::MatMul_18335_11_0_271" - ], - "out_args": [ - "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "1024", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "1280" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/attentions.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758", - "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.0/upsamplers.0/Resize", - "type": "SDResize_bfp", - "in_args": [ - "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222" - ], - "const_args": [ - "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0" - ], - "out_args": [ - "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "32", - "32", - "1280" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_30-/unet/up_blocks.0/upsamplers.0/conv/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0" - ], - "const_args": [ - "NhwcConv_30_weight_NHWC" - ], - "out_args": [ - "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "1280", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/Concat", - "type": "SDConcat_bfp", - "in_args": [ - "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30", - "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1920" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_30", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" - ], - "const_args": [ - "GroupNorm_30_wts_6_0_22" - ], - "out_args": [ - "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1920" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "3840" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_31-/unet/up_blocks.1/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30" - ], - "const_args": [ - "NhwcConv_31_weight_NHWC" - ], - "out_args": [ - "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "1920" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32", - "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_31", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223" - ], - "const_args": [ - "GroupNorm_31_wts_6_0_23" - ], - "out_args": [ - "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_32-/unet/up_blocks.1/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31" - ], - "const_args": [ - "NhwcConv_32_weight_NHWC" - ], - "out_args": [ - "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_33-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" - ], - "const_args": [ - "NhwcConv_33_weight_NHWC" - ], - "out_args": [ - "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1920" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "1", - "1", - "1920" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31", - "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_32", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" - ], - "const_args": [ - "GroupNorm_32_wts_6_1_8" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32" - ], - "const_args": [ - "onnx::MatMul_18343_11_0_272" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" - ], - "const_args": [ - "Attention_64_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" - ], - "const_args": [ - "Attention_64_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" - ], - "const_args": [ - "Attention_64_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_64", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64" - ], - "const_args": [ - "onnx::MatMul_18354_11_0_273" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763", - "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193" - ], - "const_args": [ - "onnx::MatMul_18355" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_SDCastBfp2Bf_/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_64", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_SDCastBf2Bfp_/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764" - ], - "const_args": [ - "onnx::MatMul_18365_11_0_274" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_64_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" - ], - "const_args": [ - "onnx::MatMul_18366_11_0_275_12_64_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_64_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" - ], - "const_args": [ - "onnx::MatMul_18366_11_0_275_12_64_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64" - ], - "const_args": [ - "onnx::MatMul_18367_11_0_276" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" - ], - "const_args": [ - "Attention_65_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" - ], - "const_args": [ - "Attention_65_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" - ], - "const_args": [ - "Attention_65_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_65", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65" - ], - "const_args": [ - "onnx::MatMul_18378_11_0_277" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196" - ], - "const_args": [ - "onnx::MatMul_18379" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_SDCastBfp2Bf_/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_65", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_SDCastBf2Bfp_/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772" - ], - "const_args": [ - "onnx::MatMul_18389_11_0_278" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" - ], - "const_args": [ - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_65_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" - ], - "const_args": [ - "onnx::MatMul_18390_11_0_279_12_65_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_65_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" - ], - "const_args": [ - "onnx::MatMul_18390_11_0_279_12_65_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65" - ], - "const_args": [ - "onnx::MatMul_18391_11_0_280" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230" - ], - "const_args": [ - "onnx::MatMul_18392_11_0_281" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776", - "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/Concat_1", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231", - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_33", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" - ], - "const_args": [ - "GroupNorm_33_wts_6_0_24" - ], - "out_args": [ - "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "2560" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_34-/unet/up_blocks.1/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33" - ], - "const_args": [ - "NhwcConv_34_weight_NHWC" - ], - "out_args": [ - "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35", - "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_34", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232" - ], - "const_args": [ - "GroupNorm_34_wts_6_0_25" - ], - "out_args": [ - "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_35-/unet/up_blocks.1/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34" - ], - "const_args": [ - "NhwcConv_35_weight_NHWC" - ], - "out_args": [ - "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_36-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" - ], - "const_args": [ - "NhwcConv_36_weight_NHWC" - ], - "out_args": [ - "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "1280" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "1", - "1", - "1280" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34", - "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_35", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" - ], - "const_args": [ - "GroupNorm_35_wts_6_1_9" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35" - ], - "const_args": [ - "onnx::MatMul_18399_11_0_282" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" - ], - "const_args": [ - "Attention_66_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" - ], - "const_args": [ - "Attention_66_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" - ], - "const_args": [ - "Attention_66_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_66", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66" - ], - "const_args": [ - "onnx::MatMul_18410_11_0_283" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781", - "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199" - ], - "const_args": [ - "onnx::MatMul_18411" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_SDCastBfp2Bf_/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_66", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_SDCastBf2Bfp_/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782" - ], - "const_args": [ - "onnx::MatMul_18421_11_0_284" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_66_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" - ], - "const_args": [ - "onnx::MatMul_18422_11_0_285_12_66_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_66_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" - ], - "const_args": [ - "onnx::MatMul_18422_11_0_285_12_66_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66" - ], - "const_args": [ - "onnx::MatMul_18423_11_0_286" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" - ], - "const_args": [ - "Attention_67_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" - ], - "const_args": [ - "Attention_67_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" - ], - "const_args": [ - "Attention_67_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_67", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67" - ], - "const_args": [ - "onnx::MatMul_18434_11_0_287" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202" - ], - "const_args": [ - "onnx::MatMul_18435" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_SDCastBfp2Bf_/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_67", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_SDCastBf2Bfp_/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790" - ], - "const_args": [ - "onnx::MatMul_18445_11_0_288" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" - ], - "const_args": [ - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_67_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" - ], - "const_args": [ - "onnx::MatMul_18446_11_0_289_12_67_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_67_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" - ], - "const_args": [ - "onnx::MatMul_18446_11_0_289_12_67_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67" - ], - "const_args": [ - "onnx::MatMul_18447_11_0_290" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239" - ], - "const_args": [ - "onnx::MatMul_18448_11_0_291" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794", - "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/Concat_2", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240", - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "960" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_36", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" - ], - "const_args": [ - "GroupNorm_36_wts_6_0_26" - ], - "out_args": [ - "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "960" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "960" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "960" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_37-/unet/up_blocks.1/resnets.2/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36" - ], - "const_args": [ - "NhwcConv_37_weight_NHWC" - ], - "out_args": [ - "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "960" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38", - "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_37", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241" - ], - "const_args": [ - "GroupNorm_37_wts_6_0_27" - ], - "out_args": [ - "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_38-/unet/up_blocks.1/resnets.2/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37" - ], - "const_args": [ - "NhwcConv_38_weight_NHWC" - ], - "out_args": [ - "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_39-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" - ], - "const_args": [ - "NhwcConv_39_weight_NHWC" - ], - "out_args": [ - "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "1", - "1", - "960" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/resnets.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37", - "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_38", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" - ], - "const_args": [ - "GroupNorm_38_wts_6_1_10" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999974752427e-07" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/proj_in/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38" - ], - "const_args": [ - "onnx::MatMul_18455_11_0_292" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" - ], - "const_args": [ - "Attention_68_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" - ], - "const_args": [ - "Attention_68_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" - ], - "const_args": [ - "Attention_68_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_68", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68" - ], - "const_args": [ - "onnx::MatMul_18466_11_0_293" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799", - "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205" - ], - "const_args": [ - "onnx::MatMul_18467" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_SDCastBfp2Bf_/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_68", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_SDCastBf2Bfp_/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800" - ], - "const_args": [ - "onnx::MatMul_18477_11_0_294" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_68_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" - ], - "const_args": [ - "onnx::MatMul_18478_11_0_295_12_68_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_68_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" - ], - "const_args": [ - "onnx::MatMul_18478_11_0_295_12_68_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68" - ], - "const_args": [ - "onnx::MatMul_18479_11_0_296" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" - ], - "const_args": [ - "Attention_69_qkv_weight_key" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" - ], - "const_args": [ - "Attention_69_qkv_weight_query" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" - ], - "const_args": [ - "Attention_69_qkv_weight_value" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "3" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "Attention_69", - "type": "SDMHA_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "4096" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v2" - ] - }, - "is_flash_mha": { - "type": "int", - "value": [ - "1" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69" - ], - "const_args": [ - "onnx::MatMul_18490_11_0_297" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208" - ], - "const_args": [ - "onnx::MatMul_18491" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "head_num": { - "type": "int", - "value": [ - "10" - ] - }, - "trans_head": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_SDCastBfp2Bf_/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", - "type": "SDCastBfp2Bf", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "10", - "4096", - "64" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "MultiHeadAttention_69", - "type": "SDFlatMHA", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139" - ], - "attrs": { - "num_heads": { - "type": "int", - "value": [ - "10" - ] - }, - "unidirectional": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "64", - "77" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "op_version": { - "type": "str", - "value": [ - "v1.1" - ] - }, - "is_flat_mha_1_1": { - "type": "int", - "value": [ - "1" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_SDCastBf2Bfp_/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808", - "type": "SDCastBf2Bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139" - ], - "const_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808" - ], - "const_args": [ - "onnx::MatMul_18501_11_0_298" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization", - "type": "SDLayerNorm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" - ], - "const_args": [ - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "-1" - ] - }, - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "stash_type": { - "type": "int", - "value": [ - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_69_0", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" - ], - "const_args": [ - "onnx::MatMul_18502_11_0_299_12_69_0" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Gelu" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_69_1", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" - ], - "const_args": [ - "onnx::MatMul_18502_11_0_299_12_69_1" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "2560" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4", - "type": "SDMul_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69" - ], - "const_args": [ - "onnx::MatMul_18503_11_0_300" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "2560" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "2560", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/proj_out/MatMul/MatMulAddFusion", - "type": "SDGemm_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248" - ], - "const_args": [ - "onnx::MatMul_18504_11_0_301" - ], - "out_args": [ - "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "4096", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "640" - ] - }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/attentions.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812", - "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.1/upsamplers.0/Resize", - "type": "SDResize_bfp", - "in_args": [ - "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249" - ], - "const_args": [ - "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1" - ], - "out_args": [ - "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "64", - "64", - "640" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_40-/unet/up_blocks.1/upsamplers.0/conv/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1" - ], - "const_args": [ - "NhwcConv_40_weight_NHWC" - ], - "out_args": [ - "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "640", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/Concat", - "type": "SDConcat_bfp", - "in_args": [ - "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40", - "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "960" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_39", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" - ], - "const_args": [ - "GroupNorm_39_wts_6_0_28" - ], - "out_args": [ - "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "960" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1920" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "960" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "960" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_41-/unet/up_blocks.2/resnets.0/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39" - ], - "const_args": [ - "NhwcConv_41_weight_NHWC" - ], - "out_args": [ - "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "960" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.0/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42", - "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_40", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250" - ], - "const_args": [ - "GroupNorm_40_wts_6_0_29" - ], - "out_args": [ - "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_42-/unet/up_blocks.2/resnets.0/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40" - ], - "const_args": [ - "NhwcConv_42_weight_NHWC" - ], - "out_args": [ - "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_43-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" - ], - "const_args": [ - "NhwcConv_43_weight_NHWC" - ], - "out_args": [ - "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "960" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "1", - "1", - "960" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.0/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41", - "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/Concat_1", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251", - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_41", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" - ], - "const_args": [ - "GroupNorm_41_wts_6_0_30" - ], - "out_args": [ - "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_44-/unet/up_blocks.2/resnets.1/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41" - ], - "const_args": [ - "NhwcConv_44_weight_NHWC" - ], - "out_args": [ - "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.1/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45", - "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_42", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252" - ], - "const_args": [ - "GroupNorm_42_wts_6_0_31" - ], - "out_args": [ - "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_45-/unet/up_blocks.2/resnets.1/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42" - ], - "const_args": [ - "NhwcConv_45_weight_NHWC" - ], - "out_args": [ - "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_46-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" - ], - "const_args": [ - "NhwcConv_46_weight_NHWC" - ], - "out_args": [ - "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "1", - "1", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.1/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44", - "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/Concat_2", - "type": "SDConcat_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253", - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" - ], - "attrs": { - "axis": { - "type": "int", - "value": [ - "3" - ] - }, - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_43", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" - ], - "const_args": [ - "GroupNorm_43_wts_6_0_32" - ], - "out_args": [ - "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "1280" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_47-/unet/up_blocks.2/resnets.2/conv1/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43" - ], - "const_args": [ - "NhwcConv_47_weight_NHWC" - ], - "out_args": [ - "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.2/Add", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48", - "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "1", - "1", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_44", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254" - ], - "const_args": [ - "GroupNorm_44_wts_6_0_33" - ], - "out_args": [ - "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_48-/unet/up_blocks.2/resnets.2/conv2/Conv", - "type": "SDConv_bfp", - "in_args": [ - "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44" - ], - "const_args": [ - "NhwcConv_48_weight_NHWC" - ], - "out_args": [ - "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_49-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv", - "type": "SDConv_bfp", - "in_args": [ - "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" - ], - "const_args": [ - "NhwcConv_49_weight_NHWC" - ], - "out_args": [ - "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "pads": { - "type": "int", - "value": [ - "0", - "0", - "0", - "0" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "640" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "320", - "1", - "1", - "640" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "/unet/up_blocks.2/resnets.2/Add_1", - "type": "SDAdd_bfp", - "in_args": [ - "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47", - "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49" - ], - "const_args": [], - "out_args": [ - "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255" - ], - "attrs": { - "a_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "b_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "c_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "is_bias_add": { - "type": "int", - "value": [ - "0" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_45", - "type": "SDGroupNorm_bfp", - "in_args": [ - "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255" - ], - "const_args": [ - "GroupNorm_45_wts_6_0_34" - ], - "out_args": [ - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45" - ], - "attrs": { - "epsilon": { - "type": "float", - "value": [ - "9.999999747378752e-06" - ] - }, - "groups": { - "type": "int", - "value": [ - "32" - ] - }, - "channels_last": { - "type": "int", - "value": [ - "1" - ] - }, - "activation": { - "type": "int", - "value": [ - "0" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "wts_shape": { - "type": "int", - "value": [ - "640" - ] - }, - "gamma_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "beta_shape": { - "type": "int", - "value": [ - "320" - ] - }, - "nonlinear": { - "type": "str", - "value": [ - "Silu" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8", - "bfloat16" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_SDCastBfp2Bf_GroupNorm_45.out7_45.out6_0_34", - "type": "SDCastBfp2Bf", - "in_args": [ - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45" - ], - "const_args": [ - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts" - ], - "out_args": [ - "GroupNorm_45.out7_45.out6_0_34" - ], - "attrs": { - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfp16ebs8" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - }, - { - "name": "NhwcConv_50-/unet/conv_out/Conv", - "type": "SDConv", - "in_args": [ - "GroupNorm_45.out7_45.out6_0_34" - ], - "const_args": [ - "NhwcConv_50_weight_NHWC" - ], - "out_args": [ - "NhwcConv_50_out-out_sample.out0_1_50" - ], - "attrs": { - "auto_pad": { - "type": "str", - "value": [ - "NOTSET" - ] - }, - "dilations": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "group": { - "type": "int", - "value": [ - "1" - ] - }, - "kernel_shape": { - "type": "int", - "value": [ - "3", - "3" - ] - }, - "pads": { - "type": "int", - "value": [ - "1", - "1", - "1", - "1" - ] - }, - "strides": { - "type": "int", - "value": [ - "1", - "1" - ] - }, - "input_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "320" - ] - }, - "output_shape": { - "type": "int", - "value": [ - "2", - "128", - "128", - "4" - ] - }, - "weight_shape": { - "type": "int", - "value": [ - "4", - "3", - "3", - "320" - ] - }, - "out_dtypes": { - "type": "str", - "value": [ - "bfloat16" - ] - }, - "in_dtypes": { - "type": "str", - "value": [ - "bfloat16", - "bfp16ebs8", - "float" - ] - }, - "ctrl_packet": { - "type": "int", - "value": [ - "1" - ] - } - } - } - ], - "fused_tensors": { - "in": { - "buffer_size": 88584448, - "xrt_arg_id": 0, - "packed_tensors": [ - "Transpose_34_out-sample.out0_1_0", - "/unet/Concat_3_output_0.out2_3_1", - "encoder_hidden_states.out2_6_0", - "/unet/time_proj/Concat_1_output_0.out2_3_0" - ] - }, - "out": { - "buffer_size": 262144, - "xrt_arg_id": 1, - "packed_tensors": [ - "NhwcConv_50_out-out_sample.out0_1_50" - ] - }, - "scratch": { - "buffer_size": 7679737600, - "xrt_arg_id": 2, - "packed_tensors": [ - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0", - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", - "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0", - "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1", - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1", - "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67", - "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0", - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0", - "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2", - "/unet/Add_output_0.out_17_1_0", - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2", - "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0", - "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1", - "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1", - "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1", - "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2", - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2", - "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2", - "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1", - "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3", - "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3", - "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3", - "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4", - "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4", - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5", - "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4", - "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2", - "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7", - "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5", - "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5", - "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3", - "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4", - "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5", - "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6", - "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127", - "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8", - "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9", - "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139", - "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11", - "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12", - "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13", - "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14", - "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15", - "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16", - "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8", - "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6", - "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6", - "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6", - "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12", - "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248", - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13", - "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7", - "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9", - "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14", - "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8", - "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10", - "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15", - "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9", - "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21", - "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266", - "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22", - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11", - "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10", - "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13", - "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23", - "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11", - "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14", - "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12", - "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24", - "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12", - "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54", - "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348", - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55", - "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13", - "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15", - "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56", - "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14", - "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16", - "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57", - "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15", - "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87", - "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430", - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88", - "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16", - "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17", - "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89", - "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17", - "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18", - "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90", - "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18", - "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92", - "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74", - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437", - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438", - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24", - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439", - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95", - "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77", - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445", - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446", - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25", - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447", - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98", - "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80", - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453", - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454", - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26", - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455", - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101", - "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83", - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461", - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462", - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27", - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463", - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104", - "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86", - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469", - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470", - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28", - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471", - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107", - "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89", - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477", - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478", - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29", - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479", - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110", - "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92", - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485", - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486", - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30", - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487", - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113", - "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95", - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493", - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494", - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31", - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495", - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116", - "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98", - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501", - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502", - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32", - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503", - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507", - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508", - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119", - "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101", - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509", - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510", - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33", - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511", - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120", - "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512", - "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121", - "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19", - "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19", - "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122", - "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20", - "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20", - "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123", - "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0", - "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21", - "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22", - "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124", - "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22", - "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23", - "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21", - "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125", - "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23", - "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155", - "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594", - "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156", - "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1", - "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24", - "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25", - "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157", - "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25", - "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26", - "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24", - "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158", - "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26", - "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188", - "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676", - "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189", - "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2", - "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27", - "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28", - "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190", - "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28", - "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29", - "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27", - "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191", - "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29", - "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221", - "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758", - "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222", - "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0", - "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30", - "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3", - "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30", - "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32", - "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223", - "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31", - "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33", - "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31", - "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224", - "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32", - "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230", - "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776", - "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231", - "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4", - "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33", - "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35", - "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232", - "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34", - "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36", - "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34", - "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233", - "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35", - "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239", - "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794", - "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240", - "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5", - "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36", - "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38", - "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241", - "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37", - "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39", - "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37", - "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242", - "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38", - "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248", - "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812", - "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249", - "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1", - "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40", - "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6", - "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39", - "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42", - "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250", - "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40", - "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43", - "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41", - "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251", - "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7", - "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41", - "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45", - "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252", - "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42", - "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46", - "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44", - "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253", - "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8", - "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43", - "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48", - "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254", - "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44", - "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49", - "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47", - "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255", - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45", - "GroupNorm_45.out7_45.out6_0_34" - ] - }, - "const": { - "buffer_size": 2936555456, - "xrt_arg_id": 3, - "packed_tensors": [ - "NhwcConv_0_weight_NHWC", - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts", - "GroupNorm_0_wts_6_0_0", - "unet.add_embedding.linear_1.weight_11_1_0", - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1", - "unet.add_embedding.linear_2.weight_11_1_1", - "onnx::MatMul_16751", - "onnx::MatMul_16752", - "onnx::MatMul_16775", - "onnx::MatMul_16776", - "onnx::MatMul_16807", - "onnx::MatMul_16808", - "onnx::MatMul_16831", - "onnx::MatMul_16832", - "onnx::MatMul_16863", - "onnx::MatMul_16864", - "onnx::MatMul_16887", - "onnx::MatMul_16888", - "onnx::MatMul_16911", - "onnx::MatMul_16912", - "onnx::MatMul_16935", - "onnx::MatMul_16936", - "onnx::MatMul_16959", - "onnx::MatMul_16960", - "onnx::MatMul_16983", - "onnx::MatMul_16984", - "onnx::MatMul_17007", - "onnx::MatMul_17008", - "onnx::MatMul_17031", - "onnx::MatMul_17032", - "onnx::MatMul_17055", - "onnx::MatMul_17056", - "onnx::MatMul_17079", - "onnx::MatMul_17080", - "onnx::MatMul_17111", - "onnx::MatMul_17112", - "onnx::MatMul_17135", - "onnx::MatMul_17136", - "onnx::MatMul_17159", - "onnx::MatMul_17160", - "onnx::MatMul_17183", - "onnx::MatMul_17184", - "onnx::MatMul_17207", - "onnx::MatMul_17208", - "onnx::MatMul_17231", - "onnx::MatMul_17232", - "onnx::MatMul_17255", - "onnx::MatMul_17256", - "onnx::MatMul_17279", - "onnx::MatMul_17280", - "onnx::MatMul_17303", - "onnx::MatMul_17304", - "onnx::MatMul_17327", - "onnx::MatMul_17328", - "onnx::MatMul_17359", - "onnx::MatMul_17360", - "onnx::MatMul_17383", - "onnx::MatMul_17384", - "onnx::MatMul_17407", - "onnx::MatMul_17408", - "onnx::MatMul_17431", - "onnx::MatMul_17432", - "onnx::MatMul_17455", - "onnx::MatMul_17456", - "onnx::MatMul_17479", - "onnx::MatMul_17480", - "onnx::MatMul_17503", - "onnx::MatMul_17504", - "onnx::MatMul_17527", - "onnx::MatMul_17528", - "onnx::MatMul_17551", - "onnx::MatMul_17552", - "onnx::MatMul_17575", - "onnx::MatMul_17576", - "unet.time_embedding.linear_1.weight_11_1_2", - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0", - "unet.time_embedding.linear_2.weight_11_1_3", - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2", - "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4", - "NhwcConv_1_weight_NHWC", - "GroupNorm_1_wts_6_0_1", - "NhwcConv_2_weight_NHWC", - "GroupNorm_2_wts_6_0_2", - "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5", - "NhwcConv_3_weight_NHWC", - "GroupNorm_3_wts_6_0_3", - "NhwcConv_4_weight_NHWC", - "NhwcConv_5_weight_NHWC", - "GroupNorm_4_wts_6_0_4", - "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6", - "NhwcConv_6_weight_NHWC", - "GroupNorm_5_wts_6_0_5", - "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7", - "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8", - "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9", - "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10", - "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11", - "onnx::MatMul_17611", - "onnx::MatMul_17612", - "onnx::MatMul_17635", - "onnx::MatMul_17636", - "onnx::MatMul_17659", - "onnx::MatMul_17660", - "onnx::MatMul_17683", - "onnx::MatMul_17684", - "onnx::MatMul_17707", - "onnx::MatMul_17708", - "onnx::MatMul_17731", - "onnx::MatMul_17732", - "onnx::MatMul_17755", - "onnx::MatMul_17756", - "onnx::MatMul_17779", - "onnx::MatMul_17780", - "onnx::MatMul_17803", - "onnx::MatMul_17804", - "onnx::MatMul_17827", - "onnx::MatMul_17828", - "onnx::MatMul_17859", - "onnx::MatMul_17860", - "onnx::MatMul_17883", - "onnx::MatMul_17884", - "onnx::MatMul_17907", - "onnx::MatMul_17908", - "onnx::MatMul_17931", - "onnx::MatMul_17932", - "onnx::MatMul_17955", - "onnx::MatMul_17956", - "onnx::MatMul_17979", - "onnx::MatMul_17980", - "onnx::MatMul_18003", - "onnx::MatMul_18004", - "onnx::MatMul_18027", - "onnx::MatMul_18028", - "onnx::MatMul_18051", - "onnx::MatMul_18052", - "onnx::MatMul_18075", - "onnx::MatMul_18076", - "onnx::MatMul_18107", - "onnx::MatMul_18108", - "onnx::MatMul_18131", - "onnx::MatMul_18132", - "onnx::MatMul_18155", - "onnx::MatMul_18156", - "onnx::MatMul_18179", - "onnx::MatMul_18180", - "onnx::MatMul_18203", - "onnx::MatMul_18204", - "onnx::MatMul_18227", - "onnx::MatMul_18228", - "onnx::MatMul_18251", - "onnx::MatMul_18252", - "onnx::MatMul_18275", - "onnx::MatMul_18276", - "onnx::MatMul_18299", - "onnx::MatMul_18300", - "onnx::MatMul_18323", - "onnx::MatMul_18324", - "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12", - "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13", - "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14", - "onnx::MatMul_18356", - "onnx::MatMul_18357", - "onnx::MatMul_18380", - "onnx::MatMul_18381", - "onnx::MatMul_18412", - "onnx::MatMul_18413", - "onnx::MatMul_18436", - "onnx::MatMul_18437", - "onnx::MatMul_18468", - "onnx::MatMul_18469", - "onnx::MatMul_18492", - "onnx::MatMul_18493", - "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15", - "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16", - "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17", - "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18", - "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19", - "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20", - "NhwcConv_7_weight_NHWC", - "NhwcConv_8_weight_NHWC", - "GroupNorm_6_wts_6_1_0", - "onnx::MatMul_16738_11_0_0", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", - "Attention_0_qkv_weight_key", - "Attention_0_qkv_weight_query", - "Attention_0_qkv_weight_value", - "onnx::MatMul_16749_11_0_1", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", - "onnx::MatMul_16750", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts", - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts", - "onnx::MatMul_16760_11_0_2", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", - "onnx::MatMul_16761_11_0_3_12_0_0", - "onnx::MatMul_16761_11_0_3_12_0_1", - "onnx::MatMul_16762_11_0_4", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias", - "Attention_1_qkv_weight_key", - "Attention_1_qkv_weight_query", - "Attention_1_qkv_weight_value", - "onnx::MatMul_16773_11_0_5", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias", - "onnx::MatMul_16774", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts", - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts", - "onnx::MatMul_16784_11_0_6", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias", - "onnx::MatMul_16785_11_0_7_12_1_0", - "onnx::MatMul_16785_11_0_7_12_1_1", - "onnx::MatMul_16786_11_0_8", - "onnx::MatMul_16787_11_0_9", - "GroupNorm_7_wts_6_0_6", - "NhwcConv_9_weight_NHWC", - "GroupNorm_8_wts_6_0_7", - "NhwcConv_10_weight_NHWC", - "GroupNorm_9_wts_6_1_1", - "onnx::MatMul_16794_11_0_10", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", - "Attention_2_qkv_weight_key", - "Attention_2_qkv_weight_query", - "Attention_2_qkv_weight_value", - "onnx::MatMul_16805_11_0_11", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", - "onnx::MatMul_16806", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts", - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts", - "onnx::MatMul_16816_11_0_12", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", - "onnx::MatMul_16817_11_0_13_12_2_0", - "onnx::MatMul_16817_11_0_13_12_2_1", - "onnx::MatMul_16818_11_0_14", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias", - "Attention_3_qkv_weight_key", - "Attention_3_qkv_weight_query", - "Attention_3_qkv_weight_value", - "onnx::MatMul_16829_11_0_15", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias", - "onnx::MatMul_16830", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts", - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts", - "onnx::MatMul_16840_11_0_16", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias", - "onnx::MatMul_16841_11_0_17_12_3_0", - "onnx::MatMul_16841_11_0_17_12_3_1", - "onnx::MatMul_16842_11_0_18", - "onnx::MatMul_16843_11_0_19", - "NhwcConv_11_weight_NHWC", - "GroupNorm_10_wts_6_0_8", - "NhwcConv_12_weight_NHWC", - "GroupNorm_11_wts_6_0_9", - "NhwcConv_13_weight_NHWC", - "NhwcConv_14_weight_NHWC", - "GroupNorm_12_wts_6_1_2", - "onnx::MatMul_16850_11_0_20", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias", - "Attention_4_qkv_weight_key", - "Attention_4_qkv_weight_query", - "Attention_4_qkv_weight_value", - "onnx::MatMul_16861_11_0_21", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias", - "onnx::MatMul_16862", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts", - "onnx::MatMul_16872_11_0_22", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias", - "onnx::MatMul_16873_11_0_23_12_4_0", - "onnx::MatMul_16873_11_0_23_12_4_1", - "onnx::MatMul_16874_11_0_24", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias", - "Attention_5_qkv_weight_key", - "Attention_5_qkv_weight_query", - "Attention_5_qkv_weight_value", - "onnx::MatMul_16885_11_0_25", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias", - "onnx::MatMul_16886", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts", - "onnx::MatMul_16896_11_0_26", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias", - "onnx::MatMul_16897_11_0_27_12_5_0", - "onnx::MatMul_16897_11_0_27_12_5_1", - "onnx::MatMul_16898_11_0_28", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias", - "Attention_6_qkv_weight_key", - "Attention_6_qkv_weight_query", - "Attention_6_qkv_weight_value", - "onnx::MatMul_16909_11_0_29", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias", - "onnx::MatMul_16910", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts", - "onnx::MatMul_16920_11_0_30", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias", - "onnx::MatMul_16921_11_0_31_12_6_0", - "onnx::MatMul_16921_11_0_31_12_6_1", - "onnx::MatMul_16922_11_0_32", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias", - "Attention_7_qkv_weight_key", - "Attention_7_qkv_weight_query", - "Attention_7_qkv_weight_value", - "onnx::MatMul_16933_11_0_33", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias", - "onnx::MatMul_16934", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts", - "onnx::MatMul_16944_11_0_34", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias", - "onnx::MatMul_16945_11_0_35_12_7_0", - "onnx::MatMul_16945_11_0_35_12_7_1", - "onnx::MatMul_16946_11_0_36", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias", - "Attention_8_qkv_weight_key", - "Attention_8_qkv_weight_query", - "Attention_8_qkv_weight_value", - "onnx::MatMul_16957_11_0_37", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias", - "onnx::MatMul_16958", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts", - "onnx::MatMul_16968_11_0_38", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias", - "onnx::MatMul_16969_11_0_39_12_8_0", - "onnx::MatMul_16969_11_0_39_12_8_1", - "onnx::MatMul_16970_11_0_40", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias", - "Attention_9_qkv_weight_key", - "Attention_9_qkv_weight_query", - "Attention_9_qkv_weight_value", - "onnx::MatMul_16981_11_0_41", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias", - "onnx::MatMul_16982", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts", - "onnx::MatMul_16992_11_0_42", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias", - "onnx::MatMul_16993_11_0_43_12_9_0", - "onnx::MatMul_16993_11_0_43_12_9_1", - "onnx::MatMul_16994_11_0_44", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias", - "Attention_10_qkv_weight_key", - "Attention_10_qkv_weight_query", - "Attention_10_qkv_weight_value", - "onnx::MatMul_17005_11_0_45", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias", - "onnx::MatMul_17006", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts", - "onnx::MatMul_17016_11_0_46", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias", - "onnx::MatMul_17017_11_0_47_12_10_0", - "onnx::MatMul_17017_11_0_47_12_10_1", - "onnx::MatMul_17018_11_0_48", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias", - "Attention_11_qkv_weight_key", - "Attention_11_qkv_weight_query", - "Attention_11_qkv_weight_value", - "onnx::MatMul_17029_11_0_49", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias", - "onnx::MatMul_17030", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts", - "onnx::MatMul_17040_11_0_50", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias", - "onnx::MatMul_17041_11_0_51_12_11_0", - "onnx::MatMul_17041_11_0_51_12_11_1", - "onnx::MatMul_17042_11_0_52", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias", - "Attention_12_qkv_weight_key", - "Attention_12_qkv_weight_query", - "Attention_12_qkv_weight_value", - "onnx::MatMul_17053_11_0_53", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias", - "onnx::MatMul_17054", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts", - "onnx::MatMul_17064_11_0_54", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias", - "onnx::MatMul_17065_11_0_55_12_12_0", - "onnx::MatMul_17065_11_0_55_12_12_1", - "onnx::MatMul_17066_11_0_56", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias", - "Attention_13_qkv_weight_key", - "Attention_13_qkv_weight_query", - "Attention_13_qkv_weight_value", - "onnx::MatMul_17077_11_0_57", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias", - "onnx::MatMul_17078", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts", - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts", - "onnx::MatMul_17088_11_0_58", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight", - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias", - "onnx::MatMul_17089_11_0_59_12_13_0", - "onnx::MatMul_17089_11_0_59_12_13_1", - "onnx::MatMul_17090_11_0_60", - "onnx::MatMul_17091_11_0_61", - "GroupNorm_13_wts_6_0_10", - "NhwcConv_15_weight_NHWC", - "GroupNorm_14_wts_6_0_11", - "NhwcConv_16_weight_NHWC", - "GroupNorm_15_wts_6_1_3", - "onnx::MatMul_17098_11_0_62", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias", - "Attention_14_qkv_weight_key", - "Attention_14_qkv_weight_query", - "Attention_14_qkv_weight_value", - "onnx::MatMul_17109_11_0_63", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias", - "onnx::MatMul_17110", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts", - "onnx::MatMul_17120_11_0_64", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias", - "onnx::MatMul_17121_11_0_65_12_14_0", - "onnx::MatMul_17121_11_0_65_12_14_1", - "onnx::MatMul_17122_11_0_66", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias", - "Attention_15_qkv_weight_key", - "Attention_15_qkv_weight_query", - "Attention_15_qkv_weight_value", - "onnx::MatMul_17133_11_0_67", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias", - "onnx::MatMul_17134", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts", - "onnx::MatMul_17144_11_0_68", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias", - "onnx::MatMul_17145_11_0_69_12_15_0", - "onnx::MatMul_17145_11_0_69_12_15_1", - "onnx::MatMul_17146_11_0_70", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias", - "Attention_16_qkv_weight_key", - "Attention_16_qkv_weight_query", - "Attention_16_qkv_weight_value", - "onnx::MatMul_17157_11_0_71", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias", - "onnx::MatMul_17158", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts", - "onnx::MatMul_17168_11_0_72", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias", - "onnx::MatMul_17169_11_0_73_12_16_0", - "onnx::MatMul_17169_11_0_73_12_16_1", - "onnx::MatMul_17170_11_0_74", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias", - "Attention_17_qkv_weight_key", - "Attention_17_qkv_weight_query", - "Attention_17_qkv_weight_value", - "onnx::MatMul_17181_11_0_75", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias", - "onnx::MatMul_17182", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts", - "onnx::MatMul_17192_11_0_76", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias", - "onnx::MatMul_17193_11_0_77_12_17_0", - "onnx::MatMul_17193_11_0_77_12_17_1", - "onnx::MatMul_17194_11_0_78", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias", - "Attention_18_qkv_weight_key", - "Attention_18_qkv_weight_query", - "Attention_18_qkv_weight_value", - "onnx::MatMul_17205_11_0_79", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias", - "onnx::MatMul_17206", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts", - "onnx::MatMul_17216_11_0_80", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias", - "onnx::MatMul_17217_11_0_81_12_18_0", - "onnx::MatMul_17217_11_0_81_12_18_1", - "onnx::MatMul_17218_11_0_82", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias", - "Attention_19_qkv_weight_key", - "Attention_19_qkv_weight_query", - "Attention_19_qkv_weight_value", - "onnx::MatMul_17229_11_0_83", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias", - "onnx::MatMul_17230", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts", - "onnx::MatMul_17240_11_0_84", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias", - "onnx::MatMul_17241_11_0_85_12_19_0", - "onnx::MatMul_17241_11_0_85_12_19_1", - "onnx::MatMul_17242_11_0_86", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias", - "Attention_20_qkv_weight_key", - "Attention_20_qkv_weight_query", - "Attention_20_qkv_weight_value", - "onnx::MatMul_17253_11_0_87", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias", - "onnx::MatMul_17254", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts", - "onnx::MatMul_17264_11_0_88", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias", - "onnx::MatMul_17265_11_0_89_12_20_0", - "onnx::MatMul_17265_11_0_89_12_20_1", - "onnx::MatMul_17266_11_0_90", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias", - "Attention_21_qkv_weight_key", - "Attention_21_qkv_weight_query", - "Attention_21_qkv_weight_value", - "onnx::MatMul_17277_11_0_91", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias", - "onnx::MatMul_17278", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts", - "onnx::MatMul_17288_11_0_92", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias", - "onnx::MatMul_17289_11_0_93_12_21_0", - "onnx::MatMul_17289_11_0_93_12_21_1", - "onnx::MatMul_17290_11_0_94", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias", - "Attention_22_qkv_weight_key", - "Attention_22_qkv_weight_query", - "Attention_22_qkv_weight_value", - "onnx::MatMul_17301_11_0_95", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias", - "onnx::MatMul_17302", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts", - "onnx::MatMul_17312_11_0_96", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias", - "onnx::MatMul_17313_11_0_97_12_22_0", - "onnx::MatMul_17313_11_0_97_12_22_1", - "onnx::MatMul_17314_11_0_98", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias", - "Attention_23_qkv_weight_key", - "Attention_23_qkv_weight_query", - "Attention_23_qkv_weight_value", - "onnx::MatMul_17325_11_0_99", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias", - "onnx::MatMul_17326", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts", - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts", - "onnx::MatMul_17336_11_0_100", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight", - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias", - "onnx::MatMul_17337_11_0_101_12_23_0", - "onnx::MatMul_17337_11_0_101_12_23_1", - "onnx::MatMul_17338_11_0_102", - "onnx::MatMul_17339_11_0_103", - "GroupNorm_16_wts_6_0_12", - "NhwcConv_17_weight_NHWC", - "GroupNorm_17_wts_6_0_13", - "NhwcConv_18_weight_NHWC", - "GroupNorm_18_wts_6_1_4", - "onnx::MatMul_17346_11_0_104", - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias", - "Attention_24_qkv_weight_key", - "Attention_24_qkv_weight_query", - "Attention_24_qkv_weight_value", - "onnx::MatMul_17357_11_0_105", - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias", - "onnx::MatMul_17358", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts", - "onnx::MatMul_17368_11_0_106", - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias", - "onnx::MatMul_17369_11_0_107_12_24_0", - "onnx::MatMul_17369_11_0_107_12_24_1", - "onnx::MatMul_17370_11_0_108", - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias", - "Attention_25_qkv_weight_key", - "Attention_25_qkv_weight_query", - "Attention_25_qkv_weight_value", - "onnx::MatMul_17381_11_0_109", - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias", - "onnx::MatMul_17382", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts", - "onnx::MatMul_17392_11_0_110", - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias", - "onnx::MatMul_17393_11_0_111_12_25_0", - "onnx::MatMul_17393_11_0_111_12_25_1", - "onnx::MatMul_17394_11_0_112", - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias", - "Attention_26_qkv_weight_key", - "Attention_26_qkv_weight_query", - "Attention_26_qkv_weight_value", - "onnx::MatMul_17405_11_0_113", - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias", - "onnx::MatMul_17406", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts", - "onnx::MatMul_17416_11_0_114", - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias", - "onnx::MatMul_17417_11_0_115_12_26_0", - "onnx::MatMul_17417_11_0_115_12_26_1", - "onnx::MatMul_17418_11_0_116", - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias", - "Attention_27_qkv_weight_key", - "Attention_27_qkv_weight_query", - "Attention_27_qkv_weight_value", - "onnx::MatMul_17429_11_0_117", - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias", - "onnx::MatMul_17430", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts", - "onnx::MatMul_17440_11_0_118", - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias", - "onnx::MatMul_17441_11_0_119_12_27_0", - "onnx::MatMul_17441_11_0_119_12_27_1", - "onnx::MatMul_17442_11_0_120", - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias", - "Attention_28_qkv_weight_key", - "Attention_28_qkv_weight_query", - "Attention_28_qkv_weight_value", - "onnx::MatMul_17453_11_0_121", - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias", - "onnx::MatMul_17454", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts", - "onnx::MatMul_17464_11_0_122", - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias", - "onnx::MatMul_17465_11_0_123_12_28_0", - "onnx::MatMul_17465_11_0_123_12_28_1", - "onnx::MatMul_17466_11_0_124", - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias", - "Attention_29_qkv_weight_key", - "Attention_29_qkv_weight_query", - "Attention_29_qkv_weight_value", - "onnx::MatMul_17477_11_0_125", - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias", - "onnx::MatMul_17478", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts", - "onnx::MatMul_17488_11_0_126", - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias", - "onnx::MatMul_17489_11_0_127_12_29_0", - "onnx::MatMul_17489_11_0_127_12_29_1", - "onnx::MatMul_17490_11_0_128", - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias", - "Attention_30_qkv_weight_key", - "Attention_30_qkv_weight_query", - "Attention_30_qkv_weight_value", - "onnx::MatMul_17501_11_0_129", - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias", - "onnx::MatMul_17502", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts", - "onnx::MatMul_17512_11_0_130", - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias", - "onnx::MatMul_17513_11_0_131_12_30_0", - "onnx::MatMul_17513_11_0_131_12_30_1", - "onnx::MatMul_17514_11_0_132", - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias", - "Attention_31_qkv_weight_key", - "Attention_31_qkv_weight_query", - "Attention_31_qkv_weight_value", - "onnx::MatMul_17525_11_0_133", - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias", - "onnx::MatMul_17526", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts", - "onnx::MatMul_17536_11_0_134", - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias", - "onnx::MatMul_17537_11_0_135_12_31_0", - "onnx::MatMul_17537_11_0_135_12_31_1", - "onnx::MatMul_17538_11_0_136", - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias", - "Attention_32_qkv_weight_key", - "Attention_32_qkv_weight_query", - "Attention_32_qkv_weight_value", - "onnx::MatMul_17549_11_0_137", - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias", - "onnx::MatMul_17550", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts", - "onnx::MatMul_17560_11_0_138", - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias", - "onnx::MatMul_17561_11_0_139_12_32_0", - "onnx::MatMul_17561_11_0_139_12_32_1", - "onnx::MatMul_17562_11_0_140", - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias", - "Attention_33_qkv_weight_key", - "Attention_33_qkv_weight_query", - "Attention_33_qkv_weight_value", - "onnx::MatMul_17573_11_0_141", - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias", - "onnx::MatMul_17574", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts", - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts", - "onnx::MatMul_17584_11_0_142", - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight", - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias", - "onnx::MatMul_17585_11_0_143_12_33_0", - "onnx::MatMul_17585_11_0_143_12_33_1", - "onnx::MatMul_17586_11_0_144", - "onnx::MatMul_17587_11_0_145", - "GroupNorm_19_wts_6_0_14", - "NhwcConv_19_weight_NHWC", - "GroupNorm_20_wts_6_0_15", - "NhwcConv_20_weight_NHWC", - "GroupNorm_21_wts_6_0_16", - "NhwcConv_21_weight_NHWC", - "GroupNorm_22_wts_6_0_17", - "NhwcConv_22_weight_NHWC", - "NhwcConv_23_weight_NHWC", - "GroupNorm_23_wts_6_1_5", - "onnx::MatMul_17598_11_0_146", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias", - "Attention_34_qkv_weight_key", - "Attention_34_qkv_weight_query", - "Attention_34_qkv_weight_value", - "onnx::MatMul_17609_11_0_147", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias", - "onnx::MatMul_17610", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts", - "onnx::MatMul_17620_11_0_148", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias", - "onnx::MatMul_17621_11_0_149_12_34_0", - "onnx::MatMul_17621_11_0_149_12_34_1", - "onnx::MatMul_17622_11_0_150", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias", - "Attention_35_qkv_weight_key", - "Attention_35_qkv_weight_query", - "Attention_35_qkv_weight_value", - "onnx::MatMul_17633_11_0_151", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias", - "onnx::MatMul_17634", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts", - "onnx::MatMul_17644_11_0_152", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias", - "onnx::MatMul_17645_11_0_153_12_35_0", - "onnx::MatMul_17645_11_0_153_12_35_1", - "onnx::MatMul_17646_11_0_154", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias", - "Attention_36_qkv_weight_key", - "Attention_36_qkv_weight_query", - "Attention_36_qkv_weight_value", - "onnx::MatMul_17657_11_0_155", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias", - "onnx::MatMul_17658", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts", - "onnx::MatMul_17668_11_0_156", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias", - "onnx::MatMul_17669_11_0_157_12_36_0", - "onnx::MatMul_17669_11_0_157_12_36_1", - "onnx::MatMul_17670_11_0_158", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias", - "Attention_37_qkv_weight_key", - "Attention_37_qkv_weight_query", - "Attention_37_qkv_weight_value", - "onnx::MatMul_17681_11_0_159", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias", - "onnx::MatMul_17682", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts", - "onnx::MatMul_17692_11_0_160", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias", - "onnx::MatMul_17693_11_0_161_12_37_0", - "onnx::MatMul_17693_11_0_161_12_37_1", - "onnx::MatMul_17694_11_0_162", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias", - "Attention_38_qkv_weight_key", - "Attention_38_qkv_weight_query", - "Attention_38_qkv_weight_value", - "onnx::MatMul_17705_11_0_163", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias", - "onnx::MatMul_17706", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts", - "onnx::MatMul_17716_11_0_164", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias", - "onnx::MatMul_17717_11_0_165_12_38_0", - "onnx::MatMul_17717_11_0_165_12_38_1", - "onnx::MatMul_17718_11_0_166", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias", - "Attention_39_qkv_weight_key", - "Attention_39_qkv_weight_query", - "Attention_39_qkv_weight_value", - "onnx::MatMul_17729_11_0_167", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias", - "onnx::MatMul_17730", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts", - "onnx::MatMul_17740_11_0_168", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias", - "onnx::MatMul_17741_11_0_169_12_39_0", - "onnx::MatMul_17741_11_0_169_12_39_1", - "onnx::MatMul_17742_11_0_170", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias", - "Attention_40_qkv_weight_key", - "Attention_40_qkv_weight_query", - "Attention_40_qkv_weight_value", - "onnx::MatMul_17753_11_0_171", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias", - "onnx::MatMul_17754", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts", - "onnx::MatMul_17764_11_0_172", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias", - "onnx::MatMul_17765_11_0_173_12_40_0", - "onnx::MatMul_17765_11_0_173_12_40_1", - "onnx::MatMul_17766_11_0_174", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias", - "Attention_41_qkv_weight_key", - "Attention_41_qkv_weight_query", - "Attention_41_qkv_weight_value", - "onnx::MatMul_17777_11_0_175", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias", - "onnx::MatMul_17778", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts", - "onnx::MatMul_17788_11_0_176", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias", - "onnx::MatMul_17789_11_0_177_12_41_0", - "onnx::MatMul_17789_11_0_177_12_41_1", - "onnx::MatMul_17790_11_0_178", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias", - "Attention_42_qkv_weight_key", - "Attention_42_qkv_weight_query", - "Attention_42_qkv_weight_value", - "onnx::MatMul_17801_11_0_179", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias", - "onnx::MatMul_17802", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts", - "onnx::MatMul_17812_11_0_180", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias", - "onnx::MatMul_17813_11_0_181_12_42_0", - "onnx::MatMul_17813_11_0_181_12_42_1", - "onnx::MatMul_17814_11_0_182", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias", - "Attention_43_qkv_weight_key", - "Attention_43_qkv_weight_query", - "Attention_43_qkv_weight_value", - "onnx::MatMul_17825_11_0_183", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias", - "onnx::MatMul_17826", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts", - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts", - "onnx::MatMul_17836_11_0_184", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias", - "onnx::MatMul_17837_11_0_185_12_43_0", - "onnx::MatMul_17837_11_0_185_12_43_1", - "onnx::MatMul_17838_11_0_186", - "onnx::MatMul_17839_11_0_187", - "GroupNorm_24_wts_6_0_18", - "NhwcConv_24_weight_NHWC", - "GroupNorm_25_wts_6_0_19", - "NhwcConv_25_weight_NHWC", - "NhwcConv_26_weight_NHWC", - "GroupNorm_26_wts_6_1_6", - "onnx::MatMul_17846_11_0_188", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias", - "Attention_44_qkv_weight_key", - "Attention_44_qkv_weight_query", - "Attention_44_qkv_weight_value", - "onnx::MatMul_17857_11_0_189", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias", - "onnx::MatMul_17858", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts", - "onnx::MatMul_17868_11_0_190", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias", - "onnx::MatMul_17869_11_0_191_12_44_0", - "onnx::MatMul_17869_11_0_191_12_44_1", - "onnx::MatMul_17870_11_0_192", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias", - "Attention_45_qkv_weight_key", - "Attention_45_qkv_weight_query", - "Attention_45_qkv_weight_value", - "onnx::MatMul_17881_11_0_193", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias", - "onnx::MatMul_17882", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts", - "onnx::MatMul_17892_11_0_194", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias", - "onnx::MatMul_17893_11_0_195_12_45_0", - "onnx::MatMul_17893_11_0_195_12_45_1", - "onnx::MatMul_17894_11_0_196", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias", - "Attention_46_qkv_weight_key", - "Attention_46_qkv_weight_query", - "Attention_46_qkv_weight_value", - "onnx::MatMul_17905_11_0_197", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias", - "onnx::MatMul_17906", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts", - "onnx::MatMul_17916_11_0_198", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias", - "onnx::MatMul_17917_11_0_199_12_46_0", - "onnx::MatMul_17917_11_0_199_12_46_1", - "onnx::MatMul_17918_11_0_200", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias", - "Attention_47_qkv_weight_key", - "Attention_47_qkv_weight_query", - "Attention_47_qkv_weight_value", - "onnx::MatMul_17929_11_0_201", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias", - "onnx::MatMul_17930", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts", - "onnx::MatMul_17940_11_0_202", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias", - "onnx::MatMul_17941_11_0_203_12_47_0", - "onnx::MatMul_17941_11_0_203_12_47_1", - "onnx::MatMul_17942_11_0_204", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias", - "Attention_48_qkv_weight_key", - "Attention_48_qkv_weight_query", - "Attention_48_qkv_weight_value", - "onnx::MatMul_17953_11_0_205", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias", - "onnx::MatMul_17954", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts", - "onnx::MatMul_17964_11_0_206", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias", - "onnx::MatMul_17965_11_0_207_12_48_0", - "onnx::MatMul_17965_11_0_207_12_48_1", - "onnx::MatMul_17966_11_0_208", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias", - "Attention_49_qkv_weight_key", - "Attention_49_qkv_weight_query", - "Attention_49_qkv_weight_value", - "onnx::MatMul_17977_11_0_209", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias", - "onnx::MatMul_17978", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts", - "onnx::MatMul_17988_11_0_210", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias", - "onnx::MatMul_17989_11_0_211_12_49_0", - "onnx::MatMul_17989_11_0_211_12_49_1", - "onnx::MatMul_17990_11_0_212", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias", - "Attention_50_qkv_weight_key", - "Attention_50_qkv_weight_query", - "Attention_50_qkv_weight_value", - "onnx::MatMul_18001_11_0_213", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias", - "onnx::MatMul_18002", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts", - "onnx::MatMul_18012_11_0_214", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias", - "onnx::MatMul_18013_11_0_215_12_50_0", - "onnx::MatMul_18013_11_0_215_12_50_1", - "onnx::MatMul_18014_11_0_216", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias", - "Attention_51_qkv_weight_key", - "Attention_51_qkv_weight_query", - "Attention_51_qkv_weight_value", - "onnx::MatMul_18025_11_0_217", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias", - "onnx::MatMul_18026", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts", - "onnx::MatMul_18036_11_0_218", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias", - "onnx::MatMul_18037_11_0_219_12_51_0", - "onnx::MatMul_18037_11_0_219_12_51_1", - "onnx::MatMul_18038_11_0_220", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias", - "Attention_52_qkv_weight_key", - "Attention_52_qkv_weight_query", - "Attention_52_qkv_weight_value", - "onnx::MatMul_18049_11_0_221", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias", - "onnx::MatMul_18050", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts", - "onnx::MatMul_18060_11_0_222", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias", - "onnx::MatMul_18061_11_0_223_12_52_0", - "onnx::MatMul_18061_11_0_223_12_52_1", - "onnx::MatMul_18062_11_0_224", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias", - "Attention_53_qkv_weight_key", - "Attention_53_qkv_weight_query", - "Attention_53_qkv_weight_value", - "onnx::MatMul_18073_11_0_225", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias", - "onnx::MatMul_18074", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts", - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts", - "onnx::MatMul_18084_11_0_226", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias", - "onnx::MatMul_18085_11_0_227_12_53_0", - "onnx::MatMul_18085_11_0_227_12_53_1", - "onnx::MatMul_18086_11_0_228", - "onnx::MatMul_18087_11_0_229", - "GroupNorm_27_wts_6_0_20", - "NhwcConv_27_weight_NHWC", - "GroupNorm_28_wts_6_0_21", - "NhwcConv_28_weight_NHWC", - "NhwcConv_29_weight_NHWC", - "GroupNorm_29_wts_6_1_7", - "onnx::MatMul_18094_11_0_230", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias", - "Attention_54_qkv_weight_key", - "Attention_54_qkv_weight_query", - "Attention_54_qkv_weight_value", - "onnx::MatMul_18105_11_0_231", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias", - "onnx::MatMul_18106", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts", - "onnx::MatMul_18116_11_0_232", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias", - "onnx::MatMul_18117_11_0_233_12_54_0", - "onnx::MatMul_18117_11_0_233_12_54_1", - "onnx::MatMul_18118_11_0_234", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias", - "Attention_55_qkv_weight_key", - "Attention_55_qkv_weight_query", - "Attention_55_qkv_weight_value", - "onnx::MatMul_18129_11_0_235", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias", - "onnx::MatMul_18130", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts", - "onnx::MatMul_18140_11_0_236", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias", - "onnx::MatMul_18141_11_0_237_12_55_0", - "onnx::MatMul_18141_11_0_237_12_55_1", - "onnx::MatMul_18142_11_0_238", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias", - "Attention_56_qkv_weight_key", - "Attention_56_qkv_weight_query", - "Attention_56_qkv_weight_value", - "onnx::MatMul_18153_11_0_239", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias", - "onnx::MatMul_18154", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts", - "onnx::MatMul_18164_11_0_240", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias", - "onnx::MatMul_18165_11_0_241_12_56_0", - "onnx::MatMul_18165_11_0_241_12_56_1", - "onnx::MatMul_18166_11_0_242", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias", - "Attention_57_qkv_weight_key", - "Attention_57_qkv_weight_query", - "Attention_57_qkv_weight_value", - "onnx::MatMul_18177_11_0_243", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias", - "onnx::MatMul_18178", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts", - "onnx::MatMul_18188_11_0_244", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias", - "onnx::MatMul_18189_11_0_245_12_57_0", - "onnx::MatMul_18189_11_0_245_12_57_1", - "onnx::MatMul_18190_11_0_246", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias", - "Attention_58_qkv_weight_key", - "Attention_58_qkv_weight_query", - "Attention_58_qkv_weight_value", - "onnx::MatMul_18201_11_0_247", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias", - "onnx::MatMul_18202", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts", - "onnx::MatMul_18212_11_0_248", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias", - "onnx::MatMul_18213_11_0_249_12_58_0", - "onnx::MatMul_18213_11_0_249_12_58_1", - "onnx::MatMul_18214_11_0_250", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias", - "Attention_59_qkv_weight_key", - "Attention_59_qkv_weight_query", - "Attention_59_qkv_weight_value", - "onnx::MatMul_18225_11_0_251", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias", - "onnx::MatMul_18226", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts", - "onnx::MatMul_18236_11_0_252", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias", - "onnx::MatMul_18237_11_0_253_12_59_0", - "onnx::MatMul_18237_11_0_253_12_59_1", - "onnx::MatMul_18238_11_0_254", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias", - "Attention_60_qkv_weight_key", - "Attention_60_qkv_weight_query", - "Attention_60_qkv_weight_value", - "onnx::MatMul_18249_11_0_255", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias", - "onnx::MatMul_18250", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts", - "onnx::MatMul_18260_11_0_256", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias", - "onnx::MatMul_18261_11_0_257_12_60_0", - "onnx::MatMul_18261_11_0_257_12_60_1", - "onnx::MatMul_18262_11_0_258", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias", - "Attention_61_qkv_weight_key", - "Attention_61_qkv_weight_query", - "Attention_61_qkv_weight_value", - "onnx::MatMul_18273_11_0_259", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias", - "onnx::MatMul_18274", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts", - "onnx::MatMul_18284_11_0_260", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias", - "onnx::MatMul_18285_11_0_261_12_61_0", - "onnx::MatMul_18285_11_0_261_12_61_1", - "onnx::MatMul_18286_11_0_262", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias", - "Attention_62_qkv_weight_key", - "Attention_62_qkv_weight_query", - "Attention_62_qkv_weight_value", - "onnx::MatMul_18297_11_0_263", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias", - "onnx::MatMul_18298", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts", - "onnx::MatMul_18308_11_0_264", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias", - "onnx::MatMul_18309_11_0_265_12_62_0", - "onnx::MatMul_18309_11_0_265_12_62_1", - "onnx::MatMul_18310_11_0_266", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias", - "Attention_63_qkv_weight_key", - "Attention_63_qkv_weight_query", - "Attention_63_qkv_weight_value", - "onnx::MatMul_18321_11_0_267", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias", - "onnx::MatMul_18322", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts", - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts", - "onnx::MatMul_18332_11_0_268", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight", - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias", - "onnx::MatMul_18333_11_0_269_12_63_0", - "onnx::MatMul_18333_11_0_269_12_63_1", - "onnx::MatMul_18334_11_0_270", - "onnx::MatMul_18335_11_0_271", - "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0", - "NhwcConv_30_weight_NHWC", - "GroupNorm_30_wts_6_0_22", - "NhwcConv_31_weight_NHWC", - "GroupNorm_31_wts_6_0_23", - "NhwcConv_32_weight_NHWC", - "NhwcConv_33_weight_NHWC", - "GroupNorm_32_wts_6_1_8", - "onnx::MatMul_18343_11_0_272", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", - "Attention_64_qkv_weight_key", - "Attention_64_qkv_weight_query", - "Attention_64_qkv_weight_value", - "onnx::MatMul_18354_11_0_273", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", - "onnx::MatMul_18355", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts", - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts", - "onnx::MatMul_18365_11_0_274", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", - "onnx::MatMul_18366_11_0_275_12_64_0", - "onnx::MatMul_18366_11_0_275_12_64_1", - "onnx::MatMul_18367_11_0_276", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias", - "Attention_65_qkv_weight_key", - "Attention_65_qkv_weight_query", - "Attention_65_qkv_weight_value", - "onnx::MatMul_18378_11_0_277", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias", - "onnx::MatMul_18379", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts", - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts", - "onnx::MatMul_18389_11_0_278", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias", - "onnx::MatMul_18390_11_0_279_12_65_0", - "onnx::MatMul_18390_11_0_279_12_65_1", - "onnx::MatMul_18391_11_0_280", - "onnx::MatMul_18392_11_0_281", - "GroupNorm_33_wts_6_0_24", - "NhwcConv_34_weight_NHWC", - "GroupNorm_34_wts_6_0_25", - "NhwcConv_35_weight_NHWC", - "NhwcConv_36_weight_NHWC", - "GroupNorm_35_wts_6_1_9", - "onnx::MatMul_18399_11_0_282", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", - "Attention_66_qkv_weight_key", - "Attention_66_qkv_weight_query", - "Attention_66_qkv_weight_value", - "onnx::MatMul_18410_11_0_283", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", - "onnx::MatMul_18411", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts", - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts", - "onnx::MatMul_18421_11_0_284", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", - "onnx::MatMul_18422_11_0_285_12_66_0", - "onnx::MatMul_18422_11_0_285_12_66_1", - "onnx::MatMul_18423_11_0_286", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias", - "Attention_67_qkv_weight_key", - "Attention_67_qkv_weight_query", - "Attention_67_qkv_weight_value", - "onnx::MatMul_18434_11_0_287", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias", - "onnx::MatMul_18435", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts", - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts", - "onnx::MatMul_18445_11_0_288", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias", - "onnx::MatMul_18446_11_0_289_12_67_0", - "onnx::MatMul_18446_11_0_289_12_67_1", - "onnx::MatMul_18447_11_0_290", - "onnx::MatMul_18448_11_0_291", - "GroupNorm_36_wts_6_0_26", - "NhwcConv_37_weight_NHWC", - "GroupNorm_37_wts_6_0_27", - "NhwcConv_38_weight_NHWC", - "NhwcConv_39_weight_NHWC", - "GroupNorm_38_wts_6_1_10", - "onnx::MatMul_18455_11_0_292", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias", - "Attention_68_qkv_weight_key", - "Attention_68_qkv_weight_query", - "Attention_68_qkv_weight_value", - "onnx::MatMul_18466_11_0_293", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias", - "onnx::MatMul_18467", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts", - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts", - "onnx::MatMul_18477_11_0_294", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias", - "onnx::MatMul_18478_11_0_295_12_68_0", - "onnx::MatMul_18478_11_0_295_12_68_1", - "onnx::MatMul_18479_11_0_296", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias", - "Attention_69_qkv_weight_key", - "Attention_69_qkv_weight_query", - "Attention_69_qkv_weight_value", - "onnx::MatMul_18490_11_0_297", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias", - "onnx::MatMul_18491", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts", - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts", - "onnx::MatMul_18501_11_0_298", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight", - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias", - "onnx::MatMul_18502_11_0_299_12_69_0", - "onnx::MatMul_18502_11_0_299_12_69_1", - "onnx::MatMul_18503_11_0_300", - "onnx::MatMul_18504_11_0_301", - "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1", - "NhwcConv_40_weight_NHWC", - "GroupNorm_39_wts_6_0_28", - "NhwcConv_41_weight_NHWC", - "GroupNorm_40_wts_6_0_29", - "NhwcConv_42_weight_NHWC", - "NhwcConv_43_weight_NHWC", - "GroupNorm_41_wts_6_0_30", - "NhwcConv_44_weight_NHWC", - "GroupNorm_42_wts_6_0_31", - "NhwcConv_45_weight_NHWC", - "NhwcConv_46_weight_NHWC", - "GroupNorm_43_wts_6_0_32", - "NhwcConv_47_weight_NHWC", - "GroupNorm_44_wts_6_0_33", - "NhwcConv_48_weight_NHWC", - "NhwcConv_49_weight_NHWC", - "GroupNorm_45_wts_6_0_34", - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts", - "NhwcConv_50_weight_NHWC" - ] - }, - "super_instr": { - "buffer_size": 0, - "xrt_arg_id": 4, - "packed_tensors": [] - } - }, - "tensor_map": { - "Transpose_34_out-sample.out0_1_0": { - "packed_buffer_label": "in", - "xrt_arg_id": 0, - "dtype": "bfloat16", - "shape": [ - 2, - 128, - 128, - 4 - ], - "size_in_bytes": 262144, - "op_tensor_size": 262144, - "offset": 0 - }, - "/unet/Concat_3_output_0.out2_3_1": { - "packed_buffer_label": "in", - "xrt_arg_id": 0, - "dtype": "bfloat16", - "shape": [ - 2, - 2816 - ], - "size_in_bytes": 11264, - "op_tensor_size": 11264, - "offset": 262144 - }, - "encoder_hidden_states.out2_6_0": { - "packed_buffer_label": "in", - "xrt_arg_id": 0, - "dtype": "bfloat16", - "shape": [ - 2, - 77, - 2048 - ], - "size_in_bytes": 630784, - "op_tensor_size": 630784, - "offset": 87953664 - }, - "/unet/time_proj/Concat_1_output_0.out2_3_0": { - "packed_buffer_label": "in", - "xrt_arg_id": 0, - "dtype": "bfloat16", - "shape": [ - 2, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 43166720 - }, - "NhwcConv_50_out-out_sample.out0_1_50": { - "packed_buffer_label": "out", - "xrt_arg_id": 1, - "dtype": "bfloat16", - "shape": [ - 2, - 128, - 128, - 4 - ], - "size_in_bytes": 262144, - "op_tensor_size": 262144, - "offset": 0 - }, - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 0 - }, - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 20971520 - }, - "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 31457280 - }, - "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 41943040 - }, - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 41948160 - }, - "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 41953280 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 41958400 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 42155520 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 42352640 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 42549760 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 42746880 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 42944000 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 43141120 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 43338240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 43535360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 43929600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 44323840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 44718080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 45112320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 45506560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 45900800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 46295040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 46689280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 47083520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 47477760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 47872000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 48266240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 48660480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 49054720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 49448960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 49843200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 50237440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 50631680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 51025920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 51420160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 51814400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 52208640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 52602880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 52997120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 53391360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 53785600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 54179840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 54574080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 54968320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 55362560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 55756800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 56151040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 56545280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 56939520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 57333760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 57728000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 58122240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 58516480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 58910720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 59304960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 59699200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 60093440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 60487680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 60881920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 61276160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 61670400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 62064640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 62458880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 62853120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 63247360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 63641600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 64035840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 64430080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 64824320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 65218560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 65612800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 66007040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 66401280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 66795520 - }, - "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 67189760 - }, - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 67194880 - }, - "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 67200000 - }, - "/unet/Add_output_0.out_17_1_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 67205120 - }, - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 67210240 - }, - "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 67215360 - }, - "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 67216640 - }, - "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 77702400 - }, - "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 88188160 - }, - "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 98673920 - }, - "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 109159680 - }, - "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 119645440 - }, - "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 130131200 - }, - "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 130132480 - }, - "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 140618240 - }, - "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 151104000 - }, - "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 161589760 - }, - "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 172075520 - }, - "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 320 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 182561280 - }, - "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 320 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 185182720 - }, - "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 640 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 187804160 - }, - "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 187806720 - }, - "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 193049600 - }, - "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 198292480 - }, - "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 640 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 203535360 - }, - "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 203537920 - }, - "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 203543040 - }, - "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 203548160 - }, - "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 203553280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 203558400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 203952640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 204346880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 204741120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 205135360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 205529600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 205923840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 206318080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 206712320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 207106560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 207500800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 207895040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 208289280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 208683520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 209077760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 209472000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 209866240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 210260480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 210654720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 211048960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 211443200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 211837440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 212231680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 212625920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 213020160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 213414400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 213808640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 214202880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 214597120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 214991360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 215385600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 215779840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 216174080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 216568320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 216962560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 217356800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 217751040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 218145280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 218539520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 218933760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 219328000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 219722240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 220116480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 220510720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 220904960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 221299200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 221693440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 222087680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 222481920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 222876160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 223270400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 223664640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 224058880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 224453120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 224847360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 225241600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 225635840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 226030080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 226424320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 77, - 64 - ], - "size_in_bytes": 394240, - "op_tensor_size": 394240, - "offset": 226818560 - }, - "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 227212800 - }, - "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 227217920 - }, - "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 1280 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 227223040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 227228160 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 227425280 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 227622400 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 227819520 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 228016640 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 228213760 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 228410880 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 228608000 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 228805120 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 229002240 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 229199360 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 77, - 64 - ], - "size_in_bytes": 197120, - "op_tensor_size": 197120, - "offset": 229396480 - }, - "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 640 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 229593600 - }, - "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 640 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 229596160 - }, - "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 640 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 229598720 - }, - "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 229601280 - }, - "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 229602560 - }, - "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1, - 1, - 320 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 229603840 - }, - "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 229605120 - }, - "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 234848000 - }, - "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 240090880 - }, - "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 245333760 - }, - "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 250576640 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 255819520 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 261062400 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 266305280 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 271548160 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 276791040 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 282033920 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 287276800 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 292519680 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 297762560 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 303005440 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 313491200 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 323976960 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 329219840 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 334462720 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 339705600 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 344948480 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 365920000 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 386891520 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 407863040 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 413105920 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 418348800 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 423591680 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 428834560 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 434077440 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 439320320 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 444563200 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 449806080 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 455048960 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 460291840 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 465534720 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 476020480 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 486506240 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 491749120 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 496992000 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 502234880 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 507477760 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 528449280 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 549420800 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 570392320 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 575635200 - }, - "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 580878080 - }, - "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 586120960 - }, - "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 591363840 - }, - "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 596606720 - }, - "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 601849600 - }, - "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 607092480 - }, - "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 612335360 - }, - "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 617578240 - }, - "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 622821120 - }, - "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 628064000 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 633306880 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 638549760 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 643792640 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 649035520 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 654278400 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 659521280 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 664764160 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 670007040 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 675249920 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 680492800 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 690978560 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 701464320 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 706707200 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 711950080 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 717192960 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 722435840 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 743407360 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 764378880 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 785350400 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 790593280 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 795836160 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 801079040 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 806321920 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 811564800 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 816807680 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 822050560 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 827293440 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 832536320 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 837779200 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 843022080 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 853507840 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 863993600 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 869236480 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 874479360 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 879722240 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 884965120 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 905936640 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 926908160 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 947879680 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 953122560 - }, - "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 958365440 - }, - "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 963608320 - }, - "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 640 - ], - "size_in_bytes": 1310720, - "op_tensor_size": 1310720, - "offset": 968851200 - }, - "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 640 - ], - "size_in_bytes": 1310720, - "op_tensor_size": 1310720, - "offset": 970161920 - }, - "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 971472640 - }, - "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 974094080 - }, - "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 976715520 - }, - "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 979336960 - }, - "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 981958400 - }, - "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 984579840 - }, - "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 987201280 - }, - "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 989822720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 992444160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 995065600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 997687040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1000308480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1002929920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1005551360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1008172800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1010794240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1013415680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1016037120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1021280000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1026522880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1029144320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1031765760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1034387200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1037008640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1047494400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1057980160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1068465920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1071087360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1073708800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1076330240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1078951680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1081573120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1084194560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1086816000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1089437440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1092058880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1094680320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1097301760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1102544640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1107787520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1110408960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1113030400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1115651840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1118273280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1128759040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1139244800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1149730560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1152352000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1154973440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1157594880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1160216320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1162837760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1165459200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1168080640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1170702080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1173323520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1175944960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1178566400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1183809280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1189052160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1191673600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1194295040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1196916480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1199537920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1210023680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1220509440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1230995200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1233616640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1236238080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1238859520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1241480960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1244102400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1246723840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1249345280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1251966720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1254588160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1257209600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1259831040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1265073920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1270316800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1272938240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1275559680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1278181120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1280802560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1291288320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1301774080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1312259840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1314881280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1317502720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1320124160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1322745600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1325367040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1327988480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1330609920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1333231360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1335852800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1338474240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1341095680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1346338560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1351581440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1354202880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1356824320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1359445760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1362067200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1372552960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1383038720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1393524480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1396145920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1398767360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1401388800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1404010240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1406631680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1409253120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1411874560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1414496000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1417117440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1419738880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1422360320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1427603200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1432846080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1435467520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1438088960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1440710400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1443331840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1453817600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1464303360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1474789120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1477410560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1480032000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1482653440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1485274880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1487896320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1490517760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1493139200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1495760640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1498382080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1501003520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1503624960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1508867840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1514110720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1516732160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1519353600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1521975040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1524596480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1535082240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1545568000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1556053760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1558675200 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1561296640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1563918080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1566539520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1569160960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1571782400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1574403840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1577025280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1579646720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1582268160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1584889600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1590132480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1595375360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1597996800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1600618240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1603239680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1605861120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1616346880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1626832640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1637318400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1639939840 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1642561280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1645182720 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1647804160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1650425600 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1653047040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1655668480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1658289920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1660911360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1663532800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1666154240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1671397120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1676640000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1679261440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1681882880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1684504320 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1687125760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1697611520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1708097280 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1718583040 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1721204480 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1723825920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1726447360 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1729068800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1731690240 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1734311680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1736933120 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1739554560 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1742176000 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1744797440 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1747418880 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1752661760 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1757904640 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1760526080 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1763147520 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1765768960 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1768390400 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1778876160 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1789361920 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1799847680 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1802469120 - }, - "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1805090560 - }, - "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1807712000 - }, - "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1810333440 - }, - "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1812954880 - }, - "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1815576320 - }, - "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1818197760 - }, - "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1820819200 - }, - "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1823440640 - }, - "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1826062080 - }, - "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1828683520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1831304960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1833926400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1836547840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1839169280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1841790720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1844412160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1847033600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1849655040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1852276480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1854897920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1860140800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1865383680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1868005120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1870626560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1873248000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1875869440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1886355200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1896840960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1907326720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1909948160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1912569600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1915191040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1917812480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1920433920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1923055360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1925676800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1928298240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1930919680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1933541120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1936162560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 1941405440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1946648320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1949269760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1951891200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1954512640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1957134080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1967619840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 1978105600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1988591360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1991212800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1993834240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1996455680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 1999077120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2001698560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2004320000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2006941440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2009562880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2012184320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2014805760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2017427200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2022670080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2027912960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2030534400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2033155840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2035777280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2038398720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2048884480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2059370240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2069856000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2072477440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2075098880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2077720320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2080341760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2082963200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2085584640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2088206080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2090827520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2093448960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2096070400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2098691840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2103934720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2109177600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2111799040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2114420480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2117041920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2119663360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2130149120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2140634880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2151120640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2153742080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2156363520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2158984960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2161606400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2164227840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2166849280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2169470720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2172092160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2174713600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2177335040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2179956480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2185199360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2190442240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2193063680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2195685120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2198306560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2200928000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2211413760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2221899520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2232385280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2235006720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2237628160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2240249600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2242871040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2245492480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2248113920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2250735360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2253356800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2255978240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2258599680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2261221120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2266464000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2271706880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2274328320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2276949760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2279571200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2282192640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2292678400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2303164160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2313649920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2316271360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2318892800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2321514240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2324135680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2326757120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2329378560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2332000000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2334621440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2337242880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2339864320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2342485760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2347728640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2352971520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2355592960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2358214400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2360835840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2363457280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2373943040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2384428800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2394914560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2397536000 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2400157440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2402778880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2405400320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2408021760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2410643200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2413264640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2415886080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2418507520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2421128960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2423750400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2428993280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2434236160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2436857600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2439479040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2442100480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2444721920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2455207680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2465693440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2476179200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2478800640 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2481422080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2484043520 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2486664960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2489286400 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2491907840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2494529280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2497150720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2499772160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2502393600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2505015040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2510257920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2515500800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2518122240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2520743680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2523365120 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2525986560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2536472320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2546958080 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2557443840 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2560065280 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2562686720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2565308160 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2567929600 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2570551040 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2573172480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2575793920 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2578415360 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2581036800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2583658240 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2586279680 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2591522560 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2596765440 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2599386880 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2602008320 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2604629760 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2607251200 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2617736960 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2628222720 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2638708480 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2641329920 - }, - "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2643951360 - }, - "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2646572800 - }, - "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2649194240 - }, - "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2651815680 - }, - "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2654437120 - }, - "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2657058560 - }, - "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2659680000 - }, - "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2662301440 - }, - "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2664922880 - }, - "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2667544320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2670165760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2672787200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2675408640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2678030080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2680651520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2683272960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2685894400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2688515840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2691137280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2693758720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2699001600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2704244480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2706865920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2709487360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2712108800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2714730240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2725216000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2735701760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2746187520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2748808960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2751430400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2754051840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2756673280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2759294720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2761916160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2764537600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2767159040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2769780480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2772401920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2775023360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2780266240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2785509120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2788130560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2790752000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2793373440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2795994880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2806480640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2816966400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2827452160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2830073600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2832695040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2835316480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2837937920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2840559360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2843180800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2845802240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2848423680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2851045120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2853666560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2856288000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2861530880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2866773760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2869395200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2872016640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2874638080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2877259520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2887745280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2898231040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2908716800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2911338240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2913959680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2916581120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2919202560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2921824000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2924445440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2927066880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2929688320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2932309760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2934931200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2937552640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 2942795520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2948038400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2950659840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2953281280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2955902720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2958524160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2969009920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 2979495680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2989981440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2992602880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2995224320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 2997845760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3000467200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3003088640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3005710080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3008331520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3010952960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3013574400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3016195840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3018817280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3024060160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3029303040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3031924480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3034545920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3037167360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3039788800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3050274560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3060760320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3071246080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3073867520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3076488960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3079110400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3081731840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3084353280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3086974720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3089596160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3092217600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3094839040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3097460480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3100081920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3105324800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3110567680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3113189120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3115810560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3118432000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3121053440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3131539200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3142024960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3152510720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3155132160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3157753600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3160375040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3162996480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3165617920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3168239360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3170860800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3173482240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3176103680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3178725120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3181346560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3186589440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3191832320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3194453760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3197075200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3199696640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3202318080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3212803840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3223289600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3233775360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3236396800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3239018240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3241639680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3244261120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3246882560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3249504000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3252125440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3254746880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3257368320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3259989760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3262611200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3267854080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3273096960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3275718400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3278339840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3280961280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3283582720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3294068480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3304554240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3315040000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3317661440 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3320282880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3322904320 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3325525760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3328147200 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3330768640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3333390080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3336011520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3338632960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3341254400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3343875840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3349118720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3354361600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3356983040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3359604480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3362225920 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3364847360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3375333120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3385818880 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3396304640 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3398926080 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3401547520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3404168960 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3406790400 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3409411840 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3412033280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3414654720 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3417276160 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3419897600 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3422519040 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3425140480 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3430383360 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3435626240 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3438247680 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3440869120 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3443490560 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3446112000 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3456597760 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3467083520 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3477569280 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3480190720 - }, - "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3482812160 - }, - "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3485433600 - }, - "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3488055040 - }, - "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3490676480 - }, - "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3493297920 - }, - "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3495919360 - }, - "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3498540800 - }, - "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3501162240 - }, - "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 2560 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3503783680 - }, - "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 2560 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3509026560 - }, - "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3514269440 - }, - "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3516890880 - }, - "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3519512320 - }, - "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3522133760 - }, - "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3524755200 - }, - "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3527376640 - }, - "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3529998080 - }, - "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3532619520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3535240960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3537862400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3540483840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3543105280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3545726720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3548348160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3550969600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3553591040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3556212480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3558833920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3564076800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3569319680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3571941120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3574562560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3577184000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3579805440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3590291200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3600776960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3611262720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3613884160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3616505600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3619127040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3621748480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3624369920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3626991360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3629612800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3632234240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3634855680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3637477120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3640098560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3645341440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3650584320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3653205760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3655827200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3658448640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3661070080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3671555840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3682041600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3692527360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3695148800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3697770240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3700391680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3703013120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3705634560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3708256000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3710877440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3713498880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3716120320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3718741760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3721363200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3726606080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3731848960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3734470400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3737091840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3739713280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3742334720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3752820480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3763306240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3773792000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3776413440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3779034880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3781656320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3784277760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3786899200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3789520640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3792142080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3794763520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3797384960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3800006400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3802627840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3807870720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3813113600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3815735040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3818356480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3820977920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3823599360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3834085120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3844570880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3855056640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3857678080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3860299520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3862920960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3865542400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3868163840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3870785280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3873406720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3876028160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3878649600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3881271040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3883892480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3889135360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3894378240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3896999680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3899621120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3902242560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3904864000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3915349760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3925835520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3936321280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3938942720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3941564160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3944185600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3946807040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3949428480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3952049920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3954671360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3957292800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3959914240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3962535680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3965157120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 3970400000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3975642880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3978264320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3980885760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 3983507200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3986128640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 3996614400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4007100160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4017585920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4020207360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4022828800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4025450240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4028071680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4030693120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4033314560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4035936000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4038557440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4041178880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4043800320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4046421760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4051664640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4056907520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4059528960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4062150400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4064771840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4067393280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4077879040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4088364800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4098850560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4101472000 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4104093440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4106714880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4109336320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4111957760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4114579200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4117200640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4119822080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4122443520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4125064960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4127686400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4132929280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4138172160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4140793600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4143415040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4146036480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4148657920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4159143680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4169629440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4180115200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4182736640 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4185358080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4187979520 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4190600960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4193222400 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4195843840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4198465280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4201086720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4203708160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4206329600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4208951040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4214193920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4219436800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4222058240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4224679680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4227301120 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4229922560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4240408320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4250894080 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4261379840 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4264001280 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4266622720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4269244160 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4271865600 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4274487040 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4277108480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4279729920 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4282351360 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4284972800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4287594240 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4290215680 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4295458560 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4300701440 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4303322880 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4305944320 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4308565760 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4311187200 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4321672960 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4332158720 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4342644480 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4345265920 - }, - "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4347887360 - }, - "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4350508800 - }, - "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 2560 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4353130240 - }, - "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 2560 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4358373120 - }, - "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4363616000 - }, - "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4366237440 - }, - "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4368858880 - }, - "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4371480320 - }, - "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4374101760 - }, - "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4376723200 - }, - "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4379344640 - }, - "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4381966080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4384587520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4387208960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4389830400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4392451840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4395073280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4397694720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4400316160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4402937600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4405559040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4408180480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4413423360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4418666240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4421287680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4423909120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4426530560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4429152000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4439637760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4450123520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4460609280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4463230720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4465852160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4468473600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4471095040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4473716480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4476337920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4478959360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4481580800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4484202240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4486823680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4489445120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4494688000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4499930880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4502552320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4505173760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4507795200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4510416640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4520902400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4531388160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4541873920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4544495360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4547116800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4549738240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4552359680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4554981120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4557602560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4560224000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4562845440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4565466880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4568088320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4570709760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4575952640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4581195520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4583816960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4586438400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4589059840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4591681280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4602167040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4612652800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4623138560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4625760000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4628381440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4631002880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4633624320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4636245760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4638867200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4641488640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4644110080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4646731520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4649352960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4651974400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4657217280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4662460160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4665081600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4667703040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4670324480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4672945920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4683431680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4693917440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4704403200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4707024640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4709646080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4712267520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4714888960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4717510400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4720131840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4722753280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4725374720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4727996160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4730617600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4733239040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4738481920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4743724800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4746346240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4748967680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4751589120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4754210560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4764696320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4775182080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4785667840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4788289280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4790910720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4793532160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4796153600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4798775040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4801396480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4804017920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4806639360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4809260800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4811882240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4814503680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4819746560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4824989440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4827610880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4830232320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4832853760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4835475200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4845960960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4856446720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4866932480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4869553920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4872175360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4874796800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4877418240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4880039680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4882661120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4885282560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4887904000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4890525440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4893146880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4895768320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4901011200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4906254080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4908875520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4911496960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4914118400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4916739840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4927225600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4937711360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4948197120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4950818560 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4953440000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4956061440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4958682880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4961304320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4963925760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4966547200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4969168640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4971790080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4974411520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4977032960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 4982275840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4987518720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4990140160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4992761600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 4995383040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 4998004480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5008490240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5018976000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5029461760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5032083200 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5034704640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5037326080 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5039947520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5042568960 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5045190400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5047811840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5050433280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5053054720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5055676160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5058297600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5063540480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5068783360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5071404800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5074026240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5076647680 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5079269120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5089754880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5100240640 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5110726400 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5113347840 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5115969280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5118590720 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5121212160 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5123833600 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5126455040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5129076480 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5131697920 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5134319360 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5136940800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5139562240 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5144805120 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5150048000 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5152669440 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5155290880 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5157912320 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5160533760 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5171019520 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5181505280 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5191991040 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5194612480 - }, - "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5197233920 - }, - "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5199855360 - }, - "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1920 - ], - "size_in_bytes": 3932160, - "op_tensor_size": 3932160, - "offset": 5202476800 - }, - "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1920 - ], - "size_in_bytes": 3932160, - "op_tensor_size": 3932160, - "offset": 5206408960 - }, - "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5210341120 - }, - "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5212962560 - }, - "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5215584000 - }, - "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5218205440 - }, - "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5220826880 - }, - "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5223448320 - }, - "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5226069760 - }, - "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5228691200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5231312640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5233934080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5236555520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5239176960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5241798400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5244419840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5247041280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5249662720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5252284160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5254905600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5260148480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5265391360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5268012800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5270634240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5273255680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5275877120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5286362880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5296848640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5307334400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5309955840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5312577280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5315198720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5317820160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5320441600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5323063040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5325684480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5328305920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5330927360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5333548800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5336170240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5341413120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5346656000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5349277440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5351898880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5354520320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5357141760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5367627520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5378113280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5388599040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5391220480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5393841920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5396463360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5399084800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5401706240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5404327680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5406949120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5409570560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5412192000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5414813440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5417434880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5422677760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5427920640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5430542080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5433163520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5435784960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5438406400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5448892160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5459377920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5469863680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5472485120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5475106560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5477728000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5480349440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5482970880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5485592320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5488213760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5490835200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5493456640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5496078080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5498699520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5503942400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5509185280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5511806720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5514428160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5517049600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5519671040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5530156800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5540642560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5551128320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5553749760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5556371200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5558992640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5561614080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5564235520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5566856960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5569478400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5572099840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5574721280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5577342720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5579964160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5585207040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5590449920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5593071360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5595692800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5598314240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5600935680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5611421440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5621907200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5632392960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5635014400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5637635840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5640257280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5642878720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5645500160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5648121600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5650743040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5653364480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5655985920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5658607360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5661228800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5666471680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5671714560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5674336000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5676957440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5679578880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5682200320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5692686080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5703171840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5713657600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5716279040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5718900480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5721521920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5724143360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5726764800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5729386240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5732007680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5734629120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5737250560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5739872000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5742493440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5747736320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5752979200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5755600640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5758222080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5760843520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5763464960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5773950720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5784436480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5794922240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5797543680 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5800165120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5802786560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5805408000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5808029440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5810650880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5813272320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5815893760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5818515200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5821136640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5823758080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5829000960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5834243840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5836865280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5839486720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5842108160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5844729600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5855215360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5865701120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5876186880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5878808320 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5881429760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5884051200 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5886672640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5889294080 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5891915520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5894536960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5897158400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5899779840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5902401280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5905022720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5910265600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5915508480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5918129920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5920751360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5923372800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5925994240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5936480000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 5946965760 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5957451520 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5960072960 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5962694400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5965315840 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5967937280 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5970558720 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5973180160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5975801600 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5978423040 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5981044480 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5983665920 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 20, - 1024, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5986287360 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 5991530240 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5996773120 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 5999394560 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6002016000 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6004637440 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6007258880 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6017744640 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 5120 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6028230400 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6038716160 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 1024, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6041337600 - }, - "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6043959040 - }, - "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 32, - 32, - 1280 - ], - "size_in_bytes": 2621440, - "op_tensor_size": 2621440, - "offset": 6046580480 - }, - "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1280 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6049201920 - }, - "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1280 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6059687680 - }, - "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1920 - ], - "size_in_bytes": 15728640, - "op_tensor_size": 15728640, - "offset": 6070173440 - }, - "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1920 - ], - "size_in_bytes": 15728640, - "op_tensor_size": 15728640, - "offset": 6085902080 - }, - "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6101630720 - }, - "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6106873600 - }, - "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6112116480 - }, - "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6117359360 - }, - "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6122602240 - }, - "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6127845120 - }, - "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6133088000 - }, - "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6138330880 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6143573760 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6148816640 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6154059520 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6159302400 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6164545280 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6169788160 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6175031040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6180273920 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6185516800 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6190759680 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6201245440 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6211731200 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6216974080 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6222216960 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6227459840 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6232702720 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6253674240 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6274645760 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6295617280 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6300860160 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6306103040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6311345920 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6316588800 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6321831680 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6327074560 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6332317440 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6337560320 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6342803200 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6348046080 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6353288960 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6363774720 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6374260480 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6379503360 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6384746240 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6389989120 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6395232000 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6416203520 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6437175040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6458146560 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6463389440 - }, - "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6468632320 - }, - "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6473875200 - }, - "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1280 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6479118080 - }, - "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 1280 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6489603840 - }, - "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6500089600 - }, - "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6505332480 - }, - "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6510575360 - }, - "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6515818240 - }, - "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6521061120 - }, - "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6526304000 - }, - "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6531546880 - }, - "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6536789760 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6542032640 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6547275520 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6552518400 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6557761280 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6563004160 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6568247040 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6573489920 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6578732800 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6583975680 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6589218560 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6599704320 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6610190080 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6615432960 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6620675840 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6625918720 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6631161600 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6652133120 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6673104640 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6694076160 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6699319040 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6704561920 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6709804800 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6715047680 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6720290560 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6725533440 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6730776320 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6736019200 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6741262080 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6746504960 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6751747840 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6762233600 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6772719360 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6777962240 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6783205120 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6788448000 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6793690880 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6814662400 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 6835633920 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6856605440 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6861848320 - }, - "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6867091200 - }, - "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6872334080 - }, - "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 960 - ], - "size_in_bytes": 7864320, - "op_tensor_size": 7864320, - "offset": 6877576960 - }, - "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 960 - ], - "size_in_bytes": 7864320, - "op_tensor_size": 7864320, - "offset": 6885441280 - }, - "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6893305600 - }, - "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6898548480 - }, - "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6903791360 - }, - "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6909034240 - }, - "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6914277120 - }, - "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6919520000 - }, - "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6924762880 - }, - "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6930005760 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6935248640 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6940491520 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6945734400 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6950977280 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6956220160 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6961463040 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6966705920 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6971948800 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 6977191680 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6982434560 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 6992920320 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7003406080 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7008648960 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7013891840 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7019134720 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7024377600 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7045349120 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7066320640 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7087292160 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7092535040 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7097777920 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7103020800 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7108263680 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7113506560 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7118749440 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7123992320 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7129235200 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7134478080 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7139720960 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 10, - 4096, - 64 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7144963840 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7155449600 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7165935360 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7171178240 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7176421120 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7181664000 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7186906880 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7207878400 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 2560 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7228849920 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7249821440 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 4096, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7255064320 - }, - "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7260307200 - }, - "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 64, - 64, - 640 - ], - "size_in_bytes": 5242880, - "op_tensor_size": 5242880, - "offset": 7265550080 - }, - "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7270792960 - }, - "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7291764480 - }, - "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 960 - ], - "size_in_bytes": 31457280, - "op_tensor_size": 31457280, - "offset": 7312736000 - }, - "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 960 - ], - "size_in_bytes": 31457280, - "op_tensor_size": 31457280, - "offset": 7344193280 - }, - "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7375650560 - }, - "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7386136320 - }, - "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7396622080 - }, - "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7407107840 - }, - "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7417593600 - }, - "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7428079360 - }, - "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7438565120 - }, - "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7459536640 - }, - "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7480508160 - }, - "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7490993920 - }, - "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7501479680 - }, - "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7511965440 - }, - "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7522451200 - }, - "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7532936960 - }, - "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7543422720 - }, - "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 640 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7564394240 - }, - "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7585365760 - }, - "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7595851520 - }, - "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7606337280 - }, - "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7616823040 - }, - "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7627308800 - }, - "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7637794560 - }, - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "uint8", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 10485760, - "op_tensor_size": 10485760, - "offset": 7648280320 - }, - "GroupNorm_45.out7_45.out6_0_34": { - "packed_buffer_label": "scratch", - "xrt_arg_id": 2, - "dtype": "bfloat16", - "shape": [ - 2, - 128, - 128, - 320 - ], - "size_in_bytes": 20971520, - "op_tensor_size": 20971520, - "offset": 7658766080 - }, - "NhwcConv_0_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 31040 - ], - "size_in_bytes": 31040, - "op_tensor_size": 31040, - "offset": 0, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_0.const", - "file_size": 31040 - }, - "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 31040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1.const", - "file_size": 128 - }, - "GroupNorm_0_wts_6_0_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 31168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_2.const", - "file_size": 1280 - }, - "unet.add_embedding.linear_1.weight_11_1_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4096000 - ], - "size_in_bytes": 4096000, - "op_tensor_size": 4096000, - "offset": 32448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_3.const", - "file_size": 4096000 - }, - "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 128 - ], - "size_in_bytes": 256, - "op_tensor_size": 256, - "offset": 4128448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_4.const", - "file_size": 256 - }, - "unet.add_embedding.linear_2.weight_11_1_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 4128704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_5.const", - "file_size": 1868800 - }, - "onnx::MatMul_16751": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 5997504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_6.const", - "file_size": 1515520 - }, - "onnx::MatMul_16752": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 7513024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_7.const", - "file_size": 1515520 - }, - "onnx::MatMul_16775": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 9028544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_8.const", - "file_size": 1515520 - }, - "onnx::MatMul_16776": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 10544064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_9.const", - "file_size": 1515520 - }, - "onnx::MatMul_16807": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 12059584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_10.const", - "file_size": 1515520 - }, - "onnx::MatMul_16808": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 13575104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_11.const", - "file_size": 1515520 - }, - "onnx::MatMul_16831": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 15090624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_12.const", - "file_size": 1515520 - }, - "onnx::MatMul_16832": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 16606144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_13.const", - "file_size": 1515520 - }, - "onnx::MatMul_16863": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 18121664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_14.const", - "file_size": 2990080 - }, - "onnx::MatMul_16864": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 21111744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_15.const", - "file_size": 2990080 - }, - "onnx::MatMul_16887": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 24101824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_16.const", - "file_size": 2990080 - }, - "onnx::MatMul_16888": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 27091904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_17.const", - "file_size": 2990080 - }, - "onnx::MatMul_16911": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 30081984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_18.const", - "file_size": 2990080 - }, - "onnx::MatMul_16912": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 33072064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_19.const", - "file_size": 2990080 - }, - "onnx::MatMul_16935": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 36062144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_20.const", - "file_size": 2990080 - }, - "onnx::MatMul_16936": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 39052224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_21.const", - "file_size": 2990080 - }, - "onnx::MatMul_16959": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 42042304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_22.const", - "file_size": 2990080 - }, - "onnx::MatMul_16960": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 45032384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_23.const", - "file_size": 2990080 - }, - "onnx::MatMul_16983": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 48022464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_24.const", - "file_size": 2990080 - }, - "onnx::MatMul_16984": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 51012544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_25.const", - "file_size": 2990080 - }, - "onnx::MatMul_17007": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 54002624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_26.const", - "file_size": 2990080 - }, - "onnx::MatMul_17008": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 56992704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_27.const", - "file_size": 2990080 - }, - "onnx::MatMul_17031": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 59982784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_28.const", - "file_size": 2990080 - }, - "onnx::MatMul_17032": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 62972864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_29.const", - "file_size": 2990080 - }, - "onnx::MatMul_17055": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 65962944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_30.const", - "file_size": 2990080 - }, - "onnx::MatMul_17056": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 68953024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_31.const", - "file_size": 2990080 - }, - "onnx::MatMul_17079": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 71943104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_32.const", - "file_size": 2990080 - }, - "onnx::MatMul_17080": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 74933184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_33.const", - "file_size": 2990080 - }, - "onnx::MatMul_17111": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 77923264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_34.const", - "file_size": 2990080 - }, - "onnx::MatMul_17112": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 80913344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_35.const", - "file_size": 2990080 - }, - "onnx::MatMul_17135": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 83903424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_36.const", - "file_size": 2990080 - }, - "onnx::MatMul_17136": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 86893504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_37.const", - "file_size": 2990080 - }, - "onnx::MatMul_17159": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 89883584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_38.const", - "file_size": 2990080 - }, - "onnx::MatMul_17160": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 92873664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_39.const", - "file_size": 2990080 - }, - "onnx::MatMul_17183": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 95863744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_40.const", - "file_size": 2990080 - }, - "onnx::MatMul_17184": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 98853824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_41.const", - "file_size": 2990080 - }, - "onnx::MatMul_17207": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 101843904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_42.const", - "file_size": 2990080 - }, - "onnx::MatMul_17208": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 104833984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_43.const", - "file_size": 2990080 - }, - "onnx::MatMul_17231": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 107824064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_44.const", - "file_size": 2990080 - }, - "onnx::MatMul_17232": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 110814144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_45.const", - "file_size": 2990080 - }, - "onnx::MatMul_17255": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 113804224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_46.const", - "file_size": 2990080 - }, - "onnx::MatMul_17256": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 116794304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_47.const", - "file_size": 2990080 - }, - "onnx::MatMul_17279": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 119784384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_48.const", - "file_size": 2990080 - }, - "onnx::MatMul_17280": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 122774464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_49.const", - "file_size": 2990080 - }, - "onnx::MatMul_17303": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 125764544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_50.const", - "file_size": 2990080 - }, - "onnx::MatMul_17304": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 128754624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_51.const", - "file_size": 2990080 - }, - "onnx::MatMul_17327": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 131744704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_52.const", - "file_size": 2990080 - }, - "onnx::MatMul_17328": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 134734784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_53.const", - "file_size": 2990080 - }, - "onnx::MatMul_17359": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 137724864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_54.const", - "file_size": 2990080 - }, - "onnx::MatMul_17360": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 140714944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_55.const", - "file_size": 2990080 - }, - "onnx::MatMul_17383": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 143705024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_56.const", - "file_size": 2990080 - }, - "onnx::MatMul_17384": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 146695104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_57.const", - "file_size": 2990080 - }, - "onnx::MatMul_17407": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 149685184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_58.const", - "file_size": 2990080 - }, - "onnx::MatMul_17408": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 152675264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_59.const", - "file_size": 2990080 - }, - "onnx::MatMul_17431": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 155665344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_60.const", - "file_size": 2990080 - }, - "onnx::MatMul_17432": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 158655424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_61.const", - "file_size": 2990080 - }, - "onnx::MatMul_17455": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 161645504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_62.const", - "file_size": 2990080 - }, - "onnx::MatMul_17456": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 164635584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_63.const", - "file_size": 2990080 - }, - "onnx::MatMul_17479": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 167625664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_64.const", - "file_size": 2990080 - }, - "onnx::MatMul_17480": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 170615744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_65.const", - "file_size": 2990080 - }, - "onnx::MatMul_17503": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 173605824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_66.const", - "file_size": 2990080 - }, - "onnx::MatMul_17504": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 176595904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_67.const", - "file_size": 2990080 - }, - "onnx::MatMul_17527": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 179585984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_68.const", - "file_size": 2990080 - }, - "onnx::MatMul_17528": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 182576064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_69.const", - "file_size": 2990080 - }, - "onnx::MatMul_17551": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 185566144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_70.const", - "file_size": 2990080 - }, - "onnx::MatMul_17552": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 188556224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_71.const", - "file_size": 2990080 - }, - "onnx::MatMul_17575": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 191546304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_72.const", - "file_size": 2990080 - }, - "onnx::MatMul_17576": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 194536384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_73.const", - "file_size": 2990080 - }, - "unet.time_embedding.linear_1.weight_11_1_2": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 197526464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_74.const", - "file_size": 471040 - }, - "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 128 - ], - "size_in_bytes": 256, - "op_tensor_size": 256, - "offset": 197997504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_75.const", - "file_size": 256 - }, - "unet.time_embedding.linear_2.weight_11_1_3": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 197997760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_76.const", - "file_size": 1868800 - }, - "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 128 - ], - "size_in_bytes": 256, - "op_tensor_size": 256, - "offset": 199866560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_77.const", - "file_size": 256 - }, - "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 467200 - ], - "size_in_bytes": 467200, - "op_tensor_size": 467200, - "offset": 199866816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_78.const", - "file_size": 467200 - }, - "NhwcConv_1_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 200334016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_79.const", - "file_size": 1057280 - }, - "GroupNorm_1_wts_6_0_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 201391296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_80.const", - "file_size": 1280 - }, - "NhwcConv_2_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 201392576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_81.const", - "file_size": 1057280 - }, - "GroupNorm_2_wts_6_0_2": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 202449856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_82.const", - "file_size": 1280 - }, - "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 467200 - ], - "size_in_bytes": 467200, - "op_tensor_size": 467200, - "offset": 202451136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_83.const", - "file_size": 467200 - }, - "NhwcConv_3_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 202918336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_84.const", - "file_size": 1057280 - }, - "GroupNorm_3_wts_6_0_3": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 203975616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_85.const", - "file_size": 1280 - }, - "NhwcConv_4_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 203976896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_86.const", - "file_size": 1057280 - }, - "NhwcConv_5_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1062400 - ], - "size_in_bytes": 1062400, - "op_tensor_size": 1062400, - "offset": 205034176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_87.const", - "file_size": 1062400 - }, - "GroupNorm_4_wts_6_0_4": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 206096576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_88.const", - "file_size": 1280 - }, - "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 934400 - ], - "size_in_bytes": 934400, - "op_tensor_size": 934400, - "offset": 206097856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_89.const", - "file_size": 934400 - }, - "NhwcConv_6_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2114560 - ], - "size_in_bytes": 2114560, - "op_tensor_size": 2114560, - "offset": 207032256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_90.const", - "file_size": 2114560 - }, - "GroupNorm_5_wts_6_0_5": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 209146816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_91.const", - "file_size": 2560 - }, - "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 934400 - ], - "size_in_bytes": 934400, - "op_tensor_size": 934400, - "offset": 209149376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_92.const", - "file_size": 934400 - }, - "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 210083776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_93.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 211952576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_94.const", - "file_size": 1868800 - }, - "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 213821376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_95.const", - "file_size": 1868800 - }, - "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 215690176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_96.const", - "file_size": 1868800 - }, - "onnx::MatMul_17611": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 217558976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_97.const", - "file_size": 2990080 - }, - "onnx::MatMul_17612": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 220549056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_98.const", - "file_size": 2990080 - }, - "onnx::MatMul_17635": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 223539136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_99.const", - "file_size": 2990080 - }, - "onnx::MatMul_17636": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 226529216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_100.const", - "file_size": 2990080 - }, - "onnx::MatMul_17659": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 229519296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_101.const", - "file_size": 2990080 - }, - "onnx::MatMul_17660": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 232509376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_102.const", - "file_size": 2990080 - }, - "onnx::MatMul_17683": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 235499456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_103.const", - "file_size": 2990080 - }, - "onnx::MatMul_17684": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 238489536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_104.const", - "file_size": 2990080 - }, - "onnx::MatMul_17707": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 241479616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_105.const", - "file_size": 2990080 - }, - "onnx::MatMul_17708": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 244469696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_106.const", - "file_size": 2990080 - }, - "onnx::MatMul_17731": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 247459776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_107.const", - "file_size": 2990080 - }, - "onnx::MatMul_17732": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 250449856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_108.const", - "file_size": 2990080 - }, - "onnx::MatMul_17755": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 253439936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_109.const", - "file_size": 2990080 - }, - "onnx::MatMul_17756": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 256430016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_110.const", - "file_size": 2990080 - }, - "onnx::MatMul_17779": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 259420096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_111.const", - "file_size": 2990080 - }, - "onnx::MatMul_17780": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 262410176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_112.const", - "file_size": 2990080 - }, - "onnx::MatMul_17803": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 265400256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_113.const", - "file_size": 2990080 - }, - "onnx::MatMul_17804": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 268390336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_114.const", - "file_size": 2990080 - }, - "onnx::MatMul_17827": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 271380416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_115.const", - "file_size": 2990080 - }, - "onnx::MatMul_17828": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 274370496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_116.const", - "file_size": 2990080 - }, - "onnx::MatMul_17859": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 277360576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_117.const", - "file_size": 2990080 - }, - "onnx::MatMul_17860": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 280350656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_118.const", - "file_size": 2990080 - }, - "onnx::MatMul_17883": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 283340736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_119.const", - "file_size": 2990080 - }, - "onnx::MatMul_17884": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 286330816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_120.const", - "file_size": 2990080 - }, - "onnx::MatMul_17907": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 289320896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_121.const", - "file_size": 2990080 - }, - "onnx::MatMul_17908": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 292310976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_122.const", - "file_size": 2990080 - }, - "onnx::MatMul_17931": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 295301056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_123.const", - "file_size": 2990080 - }, - "onnx::MatMul_17932": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 298291136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_124.const", - "file_size": 2990080 - }, - "onnx::MatMul_17955": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 301281216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_125.const", - "file_size": 2990080 - }, - "onnx::MatMul_17956": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 304271296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_126.const", - "file_size": 2990080 - }, - "onnx::MatMul_17979": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 307261376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_127.const", - "file_size": 2990080 - }, - "onnx::MatMul_17980": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 310251456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_128.const", - "file_size": 2990080 - }, - "onnx::MatMul_18003": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 313241536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_129.const", - "file_size": 2990080 - }, - "onnx::MatMul_18004": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 316231616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_130.const", - "file_size": 2990080 - }, - "onnx::MatMul_18027": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 319221696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_131.const", - "file_size": 2990080 - }, - "onnx::MatMul_18028": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 322211776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_132.const", - "file_size": 2990080 - }, - "onnx::MatMul_18051": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 325201856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_133.const", - "file_size": 2990080 - }, - "onnx::MatMul_18052": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 328191936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_134.const", - "file_size": 2990080 - }, - "onnx::MatMul_18075": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 331182016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_135.const", - "file_size": 2990080 - }, - "onnx::MatMul_18076": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 334172096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_136.const", - "file_size": 2990080 - }, - "onnx::MatMul_18107": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 337162176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_137.const", - "file_size": 2990080 - }, - "onnx::MatMul_18108": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 340152256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_138.const", - "file_size": 2990080 - }, - "onnx::MatMul_18131": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 343142336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_139.const", - "file_size": 2990080 - }, - "onnx::MatMul_18132": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 346132416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_140.const", - "file_size": 2990080 - }, - "onnx::MatMul_18155": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 349122496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_141.const", - "file_size": 2990080 - }, - "onnx::MatMul_18156": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 352112576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_142.const", - "file_size": 2990080 - }, - "onnx::MatMul_18179": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 355102656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_143.const", - "file_size": 2990080 - }, - "onnx::MatMul_18180": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 358092736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_144.const", - "file_size": 2990080 - }, - "onnx::MatMul_18203": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 361082816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_145.const", - "file_size": 2990080 - }, - "onnx::MatMul_18204": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 364072896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_146.const", - "file_size": 2990080 - }, - "onnx::MatMul_18227": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 367062976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_147.const", - "file_size": 2990080 - }, - "onnx::MatMul_18228": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 370053056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_148.const", - "file_size": 2990080 - }, - "onnx::MatMul_18251": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 373043136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_149.const", - "file_size": 2990080 - }, - "onnx::MatMul_18252": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 376033216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_150.const", - "file_size": 2990080 - }, - "onnx::MatMul_18275": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 379023296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_151.const", - "file_size": 2990080 - }, - "onnx::MatMul_18276": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 382013376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_152.const", - "file_size": 2990080 - }, - "onnx::MatMul_18299": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 385003456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_153.const", - "file_size": 2990080 - }, - "onnx::MatMul_18300": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 387993536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_154.const", - "file_size": 2990080 - }, - "onnx::MatMul_18323": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 390983616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_155.const", - "file_size": 2990080 - }, - "onnx::MatMul_18324": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2990080 - ], - "size_in_bytes": 2990080, - "op_tensor_size": 2990080, - "offset": 393973696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_156.const", - "file_size": 2990080 - }, - "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 396963776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_157.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 398832576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_158.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 400701376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_159.const", - "file_size": 1868800 - }, - "onnx::MatMul_18356": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 402570176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_160.const", - "file_size": 1515520 - }, - "onnx::MatMul_18357": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 404085696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_161.const", - "file_size": 1515520 - }, - "onnx::MatMul_18380": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 405601216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_162.const", - "file_size": 1515520 - }, - "onnx::MatMul_18381": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 407116736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_163.const", - "file_size": 1515520 - }, - "onnx::MatMul_18412": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 408632256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_164.const", - "file_size": 1515520 - }, - "onnx::MatMul_18413": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 410147776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_165.const", - "file_size": 1515520 - }, - "onnx::MatMul_18436": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 411663296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_166.const", - "file_size": 1515520 - }, - "onnx::MatMul_18437": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 413178816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_167.const", - "file_size": 1515520 - }, - "onnx::MatMul_18468": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 414694336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_168.const", - "file_size": 1515520 - }, - "onnx::MatMul_18469": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 416209856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_169.const", - "file_size": 1515520 - }, - "onnx::MatMul_18492": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 417725376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_170.const", - "file_size": 1515520 - }, - "onnx::MatMul_18493": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1515520 - ], - "size_in_bytes": 1515520, - "op_tensor_size": 1515520, - "offset": 419240896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_171.const", - "file_size": 1515520 - }, - "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 934400 - ], - "size_in_bytes": 934400, - "op_tensor_size": 934400, - "offset": 420756416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_172.const", - "file_size": 934400 - }, - "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 934400 - ], - "size_in_bytes": 934400, - "op_tensor_size": 934400, - "offset": 421690816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_173.const", - "file_size": 934400 - }, - "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 934400 - ], - "size_in_bytes": 934400, - "op_tensor_size": 934400, - "offset": 422625216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_174.const", - "file_size": 934400 - }, - "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 467200 - ], - "size_in_bytes": 467200, - "op_tensor_size": 467200, - "offset": 423559616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_175.const", - "file_size": 467200 - }, - "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 467200 - ], - "size_in_bytes": 467200, - "op_tensor_size": 467200, - "offset": 424026816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_176.const", - "file_size": 467200 - }, - "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 467200 - ], - "size_in_bytes": 467200, - "op_tensor_size": 467200, - "offset": 424494016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_177.const", - "file_size": 467200 - }, - "NhwcConv_7_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 424961216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_178.const", - "file_size": 4311040 - }, - "NhwcConv_8_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 250880 - ], - "size_in_bytes": 250880, - "op_tensor_size": 250880, - "offset": 429272256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_179.const", - "file_size": 250880 - }, - "GroupNorm_6_wts_6_1_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 429523136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_180.const", - "file_size": 2560 - }, - "onnx::MatMul_16738_11_0_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 429525696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_181.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 429996736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_182.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 429998016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_183.const", - "file_size": 1280 - }, - "Attention_0_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 429999296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_184.const", - "file_size": 471040 - }, - "Attention_0_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 430470336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_185.const", - "file_size": 471040 - }, - "Attention_0_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 430941376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_186.const", - "file_size": 471040 - }, - "onnx::MatMul_16749_11_0_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 431412416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_187.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 431883456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_188.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 431884736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_189.const", - "file_size": 1280 - }, - "onnx::MatMul_16750": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 431886016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_190.const", - "file_size": 471040 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 432357056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_191.const", - "file_size": 128 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 432357184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_192.const", - "file_size": 128 - }, - "onnx::MatMul_16760_11_0_2": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 432357312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_193.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 432828352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_194.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 432829632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_195.const", - "file_size": 1280 - }, - "onnx::MatMul_16761_11_0_3_12_0_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 432830912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_196.const", - "file_size": 1868800 - }, - "onnx::MatMul_16761_11_0_3_12_0_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 434699712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_197.const", - "file_size": 1868800 - }, - "onnx::MatMul_16762_11_0_4": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 436568512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_198.const", - "file_size": 1884160 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 438452672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_199.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 438453952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_200.const", - "file_size": 1280 - }, - "Attention_1_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 438455232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_201.const", - "file_size": 471040 - }, - "Attention_1_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 438926272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_202.const", - "file_size": 471040 - }, - "Attention_1_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 439397312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_203.const", - "file_size": 471040 - }, - "onnx::MatMul_16773_11_0_5": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 439868352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_204.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 440339392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_205.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 440340672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_206.const", - "file_size": 1280 - }, - "onnx::MatMul_16774": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 440341952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_207.const", - "file_size": 471040 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 440812992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_208.const", - "file_size": 128 - }, - "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 440813120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_209.const", - "file_size": 128 - }, - "onnx::MatMul_16784_11_0_6": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 440813248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_210.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 441284288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_211.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 441285568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_212.const", - "file_size": 1280 - }, - "onnx::MatMul_16785_11_0_7_12_1_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 441286848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_213.const", - "file_size": 1868800 - }, - "onnx::MatMul_16785_11_0_7_12_1_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 443155648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_214.const", - "file_size": 1868800 - }, - "onnx::MatMul_16786_11_0_8": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 445024448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_215.const", - "file_size": 1884160 - }, - "onnx::MatMul_16787_11_0_9": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 446908608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_216.const", - "file_size": 471040 - }, - "GroupNorm_7_wts_6_0_6": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 447379648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_217.const", - "file_size": 2560 - }, - "NhwcConv_9_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 447382208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_218.const", - "file_size": 4311040 - }, - "GroupNorm_8_wts_6_0_7": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 451693248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_219.const", - "file_size": 2560 - }, - "NhwcConv_10_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 451695808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_220.const", - "file_size": 4311040 - }, - "GroupNorm_9_wts_6_1_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 456006848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_221.const", - "file_size": 2560 - }, - "onnx::MatMul_16794_11_0_10": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 456009408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_222.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 456480448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_223.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 456481728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_224.const", - "file_size": 1280 - }, - "Attention_2_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 456483008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_225.const", - "file_size": 471040 - }, - "Attention_2_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 456954048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_226.const", - "file_size": 471040 - }, - "Attention_2_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 457425088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_227.const", - "file_size": 471040 - }, - "onnx::MatMul_16805_11_0_11": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 457896128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_228.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 458367168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_229.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 458368448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_230.const", - "file_size": 1280 - }, - "onnx::MatMul_16806": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 458369728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_231.const", - "file_size": 471040 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 458840768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_232.const", - "file_size": 128 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 458840896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_233.const", - "file_size": 128 - }, - "onnx::MatMul_16816_11_0_12": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 458841024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_234.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 459312064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_235.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 459313344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_236.const", - "file_size": 1280 - }, - "onnx::MatMul_16817_11_0_13_12_2_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 459314624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_237.const", - "file_size": 1868800 - }, - "onnx::MatMul_16817_11_0_13_12_2_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 461183424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_238.const", - "file_size": 1868800 - }, - "onnx::MatMul_16818_11_0_14": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 463052224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_239.const", - "file_size": 1884160 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 464936384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_240.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 464937664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_241.const", - "file_size": 1280 - }, - "Attention_3_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 464938944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_242.const", - "file_size": 471040 - }, - "Attention_3_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 465409984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_243.const", - "file_size": 471040 - }, - "Attention_3_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 465881024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_244.const", - "file_size": 471040 - }, - "onnx::MatMul_16829_11_0_15": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 466352064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_245.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 466823104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_246.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 466824384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_247.const", - "file_size": 1280 - }, - "onnx::MatMul_16830": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 466825664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_248.const", - "file_size": 471040 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 467296704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_249.const", - "file_size": 128 - }, - "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 467296832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_250.const", - "file_size": 128 - }, - "onnx::MatMul_16840_11_0_16": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 467296960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_251.const", - "file_size": 471040 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 467768000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_252.const", - "file_size": 1280 - }, - "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 467769280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_253.const", - "file_size": 1280 - }, - "onnx::MatMul_16841_11_0_17_12_3_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 467770560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_254.const", - "file_size": 1868800 - }, - "onnx::MatMul_16841_11_0_17_12_3_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 469639360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_255.const", - "file_size": 1868800 - }, - "onnx::MatMul_16842_11_0_18": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 471508160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_256.const", - "file_size": 1884160 - }, - "onnx::MatMul_16843_11_0_19": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 473392320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_257.const", - "file_size": 471040 - }, - "NhwcConv_11_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4249600 - ], - "size_in_bytes": 4249600, - "op_tensor_size": 4249600, - "offset": 473863360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_258.const", - "file_size": 4249600 - }, - "GroupNorm_10_wts_6_0_8": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 478112960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_259.const", - "file_size": 2560 - }, - "NhwcConv_12_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 8458240 - ], - "size_in_bytes": 8458240, - "op_tensor_size": 8458240, - "offset": 478115520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_260.const", - "file_size": 8458240 - }, - "GroupNorm_11_wts_6_0_9": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 486573760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_261.const", - "file_size": 5120 - }, - "NhwcConv_13_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 486578880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_262.const", - "file_size": 17244160 - }, - "NhwcConv_14_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1003520 - ], - "size_in_bytes": 1003520, - "op_tensor_size": 1003520, - "offset": 503823040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_263.const", - "file_size": 1003520 - }, - "GroupNorm_12_wts_6_1_2": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 504826560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_264.const", - "file_size": 5120 - }, - "onnx::MatMul_16850_11_0_20": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 504831680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_265.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 506700480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_266.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 506703040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_267.const", - "file_size": 2560 - }, - "Attention_4_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 506705600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_268.const", - "file_size": 1868800 - }, - "Attention_4_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 508574400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_269.const", - "file_size": 1868800 - }, - "Attention_4_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 510443200, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_270.const", - "file_size": 1868800 - }, - "onnx::MatMul_16861_11_0_21": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 512312000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_271.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 514180800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_272.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 514183360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_273.const", - "file_size": 2560 - }, - "onnx::MatMul_16862": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 514185920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_274.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 516054720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_275.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 516054848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_276.const", - "file_size": 128 - }, - "onnx::MatMul_16872_11_0_22": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 516054976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_277.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 517923776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_278.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 517926336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_279.const", - "file_size": 2560 - }, - "onnx::MatMul_16873_11_0_23_12_4_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 517928896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_280.const", - "file_size": 7475200 - }, - "onnx::MatMul_16873_11_0_23_12_4_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 525404096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_281.const", - "file_size": 7475200 - }, - "onnx::MatMul_16874_11_0_24": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 532879296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_282.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 540354496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_283.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 540357056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_284.const", - "file_size": 2560 - }, - "Attention_5_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 540359616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_285.const", - "file_size": 1868800 - }, - "Attention_5_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 542228416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_286.const", - "file_size": 1868800 - }, - "Attention_5_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 544097216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_287.const", - "file_size": 1868800 - }, - "onnx::MatMul_16885_11_0_25": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 545966016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_288.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 547834816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_289.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 547837376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_290.const", - "file_size": 2560 - }, - "onnx::MatMul_16886": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 547839936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_291.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 549708736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_292.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 549708864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_293.const", - "file_size": 128 - }, - "onnx::MatMul_16896_11_0_26": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 549708992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_294.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 551577792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_295.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 551580352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_296.const", - "file_size": 2560 - }, - "onnx::MatMul_16897_11_0_27_12_5_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 551582912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_297.const", - "file_size": 7475200 - }, - "onnx::MatMul_16897_11_0_27_12_5_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 559058112, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_298.const", - "file_size": 7475200 - }, - "onnx::MatMul_16898_11_0_28": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 566533312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_299.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 574008512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_300.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 574011072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_301.const", - "file_size": 2560 - }, - "Attention_6_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 574013632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_302.const", - "file_size": 1868800 - }, - "Attention_6_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 575882432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_303.const", - "file_size": 1868800 - }, - "Attention_6_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 577751232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_304.const", - "file_size": 1868800 - }, - "onnx::MatMul_16909_11_0_29": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 579620032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_305.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 581488832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_306.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 581491392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_307.const", - "file_size": 2560 - }, - "onnx::MatMul_16910": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 581493952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_308.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 583362752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_309.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 583362880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_310.const", - "file_size": 128 - }, - "onnx::MatMul_16920_11_0_30": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 583363008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_311.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 585231808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_312.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 585234368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_313.const", - "file_size": 2560 - }, - "onnx::MatMul_16921_11_0_31_12_6_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 585236928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_314.const", - "file_size": 7475200 - }, - "onnx::MatMul_16921_11_0_31_12_6_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 592712128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_315.const", - "file_size": 7475200 - }, - "onnx::MatMul_16922_11_0_32": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 600187328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_316.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 607662528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_317.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 607665088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_318.const", - "file_size": 2560 - }, - "Attention_7_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 607667648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_319.const", - "file_size": 1868800 - }, - "Attention_7_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 609536448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_320.const", - "file_size": 1868800 - }, - "Attention_7_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 611405248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_321.const", - "file_size": 1868800 - }, - "onnx::MatMul_16933_11_0_33": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 613274048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_322.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 615142848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_323.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 615145408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_324.const", - "file_size": 2560 - }, - "onnx::MatMul_16934": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 615147968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_325.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 617016768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_326.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 617016896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_327.const", - "file_size": 128 - }, - "onnx::MatMul_16944_11_0_34": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 617017024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_328.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 618885824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_329.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 618888384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_330.const", - "file_size": 2560 - }, - "onnx::MatMul_16945_11_0_35_12_7_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 618890944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_331.const", - "file_size": 7475200 - }, - "onnx::MatMul_16945_11_0_35_12_7_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 626366144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_332.const", - "file_size": 7475200 - }, - "onnx::MatMul_16946_11_0_36": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 633841344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_333.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 641316544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_334.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 641319104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_335.const", - "file_size": 2560 - }, - "Attention_8_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 641321664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_336.const", - "file_size": 1868800 - }, - "Attention_8_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 643190464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_337.const", - "file_size": 1868800 - }, - "Attention_8_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 645059264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_338.const", - "file_size": 1868800 - }, - "onnx::MatMul_16957_11_0_37": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 646928064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_339.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 648796864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_340.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 648799424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_341.const", - "file_size": 2560 - }, - "onnx::MatMul_16958": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 648801984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_342.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 650670784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_343.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 650670912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_344.const", - "file_size": 128 - }, - "onnx::MatMul_16968_11_0_38": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 650671040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_345.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 652539840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_346.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 652542400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_347.const", - "file_size": 2560 - }, - "onnx::MatMul_16969_11_0_39_12_8_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 652544960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_348.const", - "file_size": 7475200 - }, - "onnx::MatMul_16969_11_0_39_12_8_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 660020160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_349.const", - "file_size": 7475200 - }, - "onnx::MatMul_16970_11_0_40": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 667495360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_350.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 674970560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_351.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 674973120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_352.const", - "file_size": 2560 - }, - "Attention_9_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 674975680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_353.const", - "file_size": 1868800 - }, - "Attention_9_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 676844480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_354.const", - "file_size": 1868800 - }, - "Attention_9_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 678713280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_355.const", - "file_size": 1868800 - }, - "onnx::MatMul_16981_11_0_41": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 680582080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_356.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 682450880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_357.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 682453440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_358.const", - "file_size": 2560 - }, - "onnx::MatMul_16982": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 682456000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_359.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 684324800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_360.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 684324928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_361.const", - "file_size": 128 - }, - "onnx::MatMul_16992_11_0_42": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 684325056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_362.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 686193856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_363.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 686196416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_364.const", - "file_size": 2560 - }, - "onnx::MatMul_16993_11_0_43_12_9_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 686198976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_365.const", - "file_size": 7475200 - }, - "onnx::MatMul_16993_11_0_43_12_9_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 693674176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_366.const", - "file_size": 7475200 - }, - "onnx::MatMul_16994_11_0_44": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 701149376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_367.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 708624576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_368.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 708627136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_369.const", - "file_size": 2560 - }, - "Attention_10_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 708629696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_370.const", - "file_size": 1868800 - }, - "Attention_10_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 710498496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_371.const", - "file_size": 1868800 - }, - "Attention_10_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 712367296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_372.const", - "file_size": 1868800 - }, - "onnx::MatMul_17005_11_0_45": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 714236096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_373.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 716104896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_374.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 716107456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_375.const", - "file_size": 2560 - }, - "onnx::MatMul_17006": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 716110016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_376.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 717978816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_377.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 717978944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_378.const", - "file_size": 128 - }, - "onnx::MatMul_17016_11_0_46": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 717979072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_379.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 719847872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_380.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 719850432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_381.const", - "file_size": 2560 - }, - "onnx::MatMul_17017_11_0_47_12_10_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 719852992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_382.const", - "file_size": 7475200 - }, - "onnx::MatMul_17017_11_0_47_12_10_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 727328192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_383.const", - "file_size": 7475200 - }, - "onnx::MatMul_17018_11_0_48": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 734803392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_384.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 742278592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_385.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 742281152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_386.const", - "file_size": 2560 - }, - "Attention_11_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 742283712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_387.const", - "file_size": 1868800 - }, - "Attention_11_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 744152512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_388.const", - "file_size": 1868800 - }, - "Attention_11_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 746021312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_389.const", - "file_size": 1868800 - }, - "onnx::MatMul_17029_11_0_49": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 747890112, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_390.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 749758912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_391.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 749761472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_392.const", - "file_size": 2560 - }, - "onnx::MatMul_17030": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 749764032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_393.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 751632832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_394.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 751632960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_395.const", - "file_size": 128 - }, - "onnx::MatMul_17040_11_0_50": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 751633088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_396.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 753501888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_397.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 753504448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_398.const", - "file_size": 2560 - }, - "onnx::MatMul_17041_11_0_51_12_11_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 753507008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_399.const", - "file_size": 7475200 - }, - "onnx::MatMul_17041_11_0_51_12_11_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 760982208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_400.const", - "file_size": 7475200 - }, - "onnx::MatMul_17042_11_0_52": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 768457408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_401.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 775932608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_402.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 775935168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_403.const", - "file_size": 2560 - }, - "Attention_12_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 775937728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_404.const", - "file_size": 1868800 - }, - "Attention_12_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 777806528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_405.const", - "file_size": 1868800 - }, - "Attention_12_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 779675328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_406.const", - "file_size": 1868800 - }, - "onnx::MatMul_17053_11_0_53": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 781544128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_407.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 783412928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_408.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 783415488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_409.const", - "file_size": 2560 - }, - "onnx::MatMul_17054": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 783418048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_410.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 785286848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_411.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 785286976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_412.const", - "file_size": 128 - }, - "onnx::MatMul_17064_11_0_54": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 785287104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_413.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 787155904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_414.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 787158464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_415.const", - "file_size": 2560 - }, - "onnx::MatMul_17065_11_0_55_12_12_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 787161024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_416.const", - "file_size": 7475200 - }, - "onnx::MatMul_17065_11_0_55_12_12_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 794636224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_417.const", - "file_size": 7475200 - }, - "onnx::MatMul_17066_11_0_56": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 802111424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_418.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 809586624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_419.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 809589184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_420.const", - "file_size": 2560 - }, - "Attention_13_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 809591744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_421.const", - "file_size": 1868800 - }, - "Attention_13_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 811460544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_422.const", - "file_size": 1868800 - }, - "Attention_13_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 813329344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_423.const", - "file_size": 1868800 - }, - "onnx::MatMul_17077_11_0_57": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 815198144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_424.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 817066944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_425.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 817069504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_426.const", - "file_size": 2560 - }, - "onnx::MatMul_17078": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 817072064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_427.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 818940864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_428.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 818940992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_429.const", - "file_size": 128 - }, - "onnx::MatMul_17088_11_0_58": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 818941120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_430.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 820809920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_431.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 820812480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_432.const", - "file_size": 2560 - }, - "onnx::MatMul_17089_11_0_59_12_13_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 820815040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_433.const", - "file_size": 7475200 - }, - "onnx::MatMul_17089_11_0_59_12_13_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 828290240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_434.const", - "file_size": 7475200 - }, - "onnx::MatMul_17090_11_0_60": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 835765440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_435.const", - "file_size": 7475200 - }, - "onnx::MatMul_17091_11_0_61": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 843240640, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_436.const", - "file_size": 1868800 - }, - "GroupNorm_13_wts_6_0_10": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 845109440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_437.const", - "file_size": 5120 - }, - "NhwcConv_15_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 845114560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_438.const", - "file_size": 17244160 - }, - "GroupNorm_14_wts_6_0_11": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 862358720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_439.const", - "file_size": 5120 - }, - "NhwcConv_16_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 862363840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_440.const", - "file_size": 17244160 - }, - "GroupNorm_15_wts_6_1_3": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 879608000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_441.const", - "file_size": 5120 - }, - "onnx::MatMul_17098_11_0_62": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 879613120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_442.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 881481920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_443.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 881484480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_444.const", - "file_size": 2560 - }, - "Attention_14_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 881487040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_445.const", - "file_size": 1868800 - }, - "Attention_14_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 883355840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_446.const", - "file_size": 1868800 - }, - "Attention_14_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 885224640, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_447.const", - "file_size": 1868800 - }, - "onnx::MatMul_17109_11_0_63": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 887093440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_448.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 888962240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_449.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 888964800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_450.const", - "file_size": 2560 - }, - "onnx::MatMul_17110": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 888967360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_451.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 890836160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_452.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 890836288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_453.const", - "file_size": 128 - }, - "onnx::MatMul_17120_11_0_64": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 890836416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_454.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 892705216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_455.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 892707776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_456.const", - "file_size": 2560 - }, - "onnx::MatMul_17121_11_0_65_12_14_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 892710336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_457.const", - "file_size": 7475200 - }, - "onnx::MatMul_17121_11_0_65_12_14_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 900185536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_458.const", - "file_size": 7475200 - }, - "onnx::MatMul_17122_11_0_66": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 907660736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_459.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 915135936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_460.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 915138496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_461.const", - "file_size": 2560 - }, - "Attention_15_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 915141056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_462.const", - "file_size": 1868800 - }, - "Attention_15_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 917009856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_463.const", - "file_size": 1868800 - }, - "Attention_15_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 918878656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_464.const", - "file_size": 1868800 - }, - "onnx::MatMul_17133_11_0_67": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 920747456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_465.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 922616256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_466.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 922618816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_467.const", - "file_size": 2560 - }, - "onnx::MatMul_17134": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 922621376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_468.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 924490176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_469.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 924490304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_470.const", - "file_size": 128 - }, - "onnx::MatMul_17144_11_0_68": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 924490432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_471.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 926359232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_472.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 926361792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_473.const", - "file_size": 2560 - }, - "onnx::MatMul_17145_11_0_69_12_15_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 926364352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_474.const", - "file_size": 7475200 - }, - "onnx::MatMul_17145_11_0_69_12_15_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 933839552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_475.const", - "file_size": 7475200 - }, - "onnx::MatMul_17146_11_0_70": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 941314752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_476.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 948789952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_477.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 948792512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_478.const", - "file_size": 2560 - }, - "Attention_16_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 948795072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_479.const", - "file_size": 1868800 - }, - "Attention_16_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 950663872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_480.const", - "file_size": 1868800 - }, - "Attention_16_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 952532672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_481.const", - "file_size": 1868800 - }, - "onnx::MatMul_17157_11_0_71": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 954401472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_482.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 956270272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_483.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 956272832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_484.const", - "file_size": 2560 - }, - "onnx::MatMul_17158": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 956275392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_485.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 958144192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_486.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 958144320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_487.const", - "file_size": 128 - }, - "onnx::MatMul_17168_11_0_72": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 958144448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_488.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 960013248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_489.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 960015808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_490.const", - "file_size": 2560 - }, - "onnx::MatMul_17169_11_0_73_12_16_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 960018368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_491.const", - "file_size": 7475200 - }, - "onnx::MatMul_17169_11_0_73_12_16_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 967493568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_492.const", - "file_size": 7475200 - }, - "onnx::MatMul_17170_11_0_74": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 974968768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_493.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 982443968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_494.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 982446528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_495.const", - "file_size": 2560 - }, - "Attention_17_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 982449088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_496.const", - "file_size": 1868800 - }, - "Attention_17_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 984317888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_497.const", - "file_size": 1868800 - }, - "Attention_17_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 986186688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_498.const", - "file_size": 1868800 - }, - "onnx::MatMul_17181_11_0_75": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 988055488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_499.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 989924288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_500.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 989926848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_501.const", - "file_size": 2560 - }, - "onnx::MatMul_17182": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 989929408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_502.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 991798208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_503.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 991798336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_504.const", - "file_size": 128 - }, - "onnx::MatMul_17192_11_0_76": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 991798464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_505.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 993667264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_506.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 993669824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_507.const", - "file_size": 2560 - }, - "onnx::MatMul_17193_11_0_77_12_17_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 993672384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_508.const", - "file_size": 7475200 - }, - "onnx::MatMul_17193_11_0_77_12_17_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1001147584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_509.const", - "file_size": 7475200 - }, - "onnx::MatMul_17194_11_0_78": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1008622784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_510.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1016097984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_511.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1016100544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_512.const", - "file_size": 2560 - }, - "Attention_18_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1016103104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_513.const", - "file_size": 1868800 - }, - "Attention_18_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1017971904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_514.const", - "file_size": 1868800 - }, - "Attention_18_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1019840704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_515.const", - "file_size": 1868800 - }, - "onnx::MatMul_17205_11_0_79": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1021709504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_516.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1023578304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_517.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1023580864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_518.const", - "file_size": 2560 - }, - "onnx::MatMul_17206": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1023583424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_519.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1025452224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_520.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1025452352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_521.const", - "file_size": 128 - }, - "onnx::MatMul_17216_11_0_80": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1025452480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_522.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1027321280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_523.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1027323840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_524.const", - "file_size": 2560 - }, - "onnx::MatMul_17217_11_0_81_12_18_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1027326400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_525.const", - "file_size": 7475200 - }, - "onnx::MatMul_17217_11_0_81_12_18_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1034801600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_526.const", - "file_size": 7475200 - }, - "onnx::MatMul_17218_11_0_82": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1042276800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_527.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1049752000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_528.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1049754560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_529.const", - "file_size": 2560 - }, - "Attention_19_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1049757120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_530.const", - "file_size": 1868800 - }, - "Attention_19_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1051625920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_531.const", - "file_size": 1868800 - }, - "Attention_19_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1053494720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_532.const", - "file_size": 1868800 - }, - "onnx::MatMul_17229_11_0_83": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1055363520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_533.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1057232320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_534.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1057234880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_535.const", - "file_size": 2560 - }, - "onnx::MatMul_17230": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1057237440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_536.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1059106240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_537.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1059106368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_538.const", - "file_size": 128 - }, - "onnx::MatMul_17240_11_0_84": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1059106496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_539.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1060975296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_540.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1060977856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_541.const", - "file_size": 2560 - }, - "onnx::MatMul_17241_11_0_85_12_19_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1060980416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_542.const", - "file_size": 7475200 - }, - "onnx::MatMul_17241_11_0_85_12_19_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1068455616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_543.const", - "file_size": 7475200 - }, - "onnx::MatMul_17242_11_0_86": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1075930816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_544.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1083406016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_545.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1083408576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_546.const", - "file_size": 2560 - }, - "Attention_20_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1083411136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_547.const", - "file_size": 1868800 - }, - "Attention_20_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1085279936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_548.const", - "file_size": 1868800 - }, - "Attention_20_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1087148736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_549.const", - "file_size": 1868800 - }, - "onnx::MatMul_17253_11_0_87": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1089017536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_550.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1090886336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_551.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1090888896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_552.const", - "file_size": 2560 - }, - "onnx::MatMul_17254": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1090891456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_553.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1092760256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_554.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1092760384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_555.const", - "file_size": 128 - }, - "onnx::MatMul_17264_11_0_88": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1092760512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_556.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1094629312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_557.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1094631872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_558.const", - "file_size": 2560 - }, - "onnx::MatMul_17265_11_0_89_12_20_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1094634432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_559.const", - "file_size": 7475200 - }, - "onnx::MatMul_17265_11_0_89_12_20_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1102109632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_560.const", - "file_size": 7475200 - }, - "onnx::MatMul_17266_11_0_90": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1109584832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_561.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1117060032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_562.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1117062592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_563.const", - "file_size": 2560 - }, - "Attention_21_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1117065152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_564.const", - "file_size": 1868800 - }, - "Attention_21_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1118933952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_565.const", - "file_size": 1868800 - }, - "Attention_21_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1120802752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_566.const", - "file_size": 1868800 - }, - "onnx::MatMul_17277_11_0_91": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1122671552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_567.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1124540352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_568.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1124542912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_569.const", - "file_size": 2560 - }, - "onnx::MatMul_17278": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1124545472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_570.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1126414272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_571.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1126414400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_572.const", - "file_size": 128 - }, - "onnx::MatMul_17288_11_0_92": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1126414528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_573.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1128283328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_574.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1128285888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_575.const", - "file_size": 2560 - }, - "onnx::MatMul_17289_11_0_93_12_21_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1128288448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_576.const", - "file_size": 7475200 - }, - "onnx::MatMul_17289_11_0_93_12_21_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1135763648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_577.const", - "file_size": 7475200 - }, - "onnx::MatMul_17290_11_0_94": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1143238848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_578.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1150714048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_579.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1150716608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_580.const", - "file_size": 2560 - }, - "Attention_22_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1150719168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_581.const", - "file_size": 1868800 - }, - "Attention_22_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1152587968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_582.const", - "file_size": 1868800 - }, - "Attention_22_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1154456768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_583.const", - "file_size": 1868800 - }, - "onnx::MatMul_17301_11_0_95": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1156325568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_584.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1158194368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_585.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1158196928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_586.const", - "file_size": 2560 - }, - "onnx::MatMul_17302": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1158199488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_587.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1160068288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_588.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1160068416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_589.const", - "file_size": 128 - }, - "onnx::MatMul_17312_11_0_96": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1160068544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_590.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1161937344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_591.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1161939904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_592.const", - "file_size": 2560 - }, - "onnx::MatMul_17313_11_0_97_12_22_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1161942464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_593.const", - "file_size": 7475200 - }, - "onnx::MatMul_17313_11_0_97_12_22_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1169417664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_594.const", - "file_size": 7475200 - }, - "onnx::MatMul_17314_11_0_98": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1176892864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_595.const", - "file_size": 7475200 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1184368064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_596.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1184370624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_597.const", - "file_size": 2560 - }, - "Attention_23_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1184373184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_598.const", - "file_size": 1868800 - }, - "Attention_23_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1186241984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_599.const", - "file_size": 1868800 - }, - "Attention_23_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1188110784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_600.const", - "file_size": 1868800 - }, - "onnx::MatMul_17325_11_0_99": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1189979584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_601.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1191848384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_602.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1191850944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_603.const", - "file_size": 2560 - }, - "onnx::MatMul_17326": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1191853504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_604.const", - "file_size": 1868800 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1193722304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_605.const", - "file_size": 128 - }, - "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1193722432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_606.const", - "file_size": 128 - }, - "onnx::MatMul_17336_11_0_100": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1193722560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_607.const", - "file_size": 1868800 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1195591360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_608.const", - "file_size": 2560 - }, - "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1195593920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_609.const", - "file_size": 2560 - }, - "onnx::MatMul_17337_11_0_101_12_23_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1195596480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_610.const", - "file_size": 7475200 - }, - "onnx::MatMul_17337_11_0_101_12_23_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1203071680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_611.const", - "file_size": 7475200 - }, - "onnx::MatMul_17338_11_0_102": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1210546880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_612.const", - "file_size": 7475200 - }, - "onnx::MatMul_17339_11_0_103": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1218022080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_613.const", - "file_size": 1868800 - }, - "GroupNorm_16_wts_6_0_12": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1219890880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_614.const", - "file_size": 5120 - }, - "NhwcConv_17_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 1219896000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_615.const", - "file_size": 17244160 - }, - "GroupNorm_17_wts_6_0_13": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1237140160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_616.const", - "file_size": 5120 - }, - "NhwcConv_18_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 1237145280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_617.const", - "file_size": 17244160 - }, - "GroupNorm_18_wts_6_1_4": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1254389440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_618.const", - "file_size": 5120 - }, - "onnx::MatMul_17346_11_0_104": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1254394560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_619.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1256263360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_620.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1256265920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_621.const", - "file_size": 2560 - }, - "Attention_24_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1256268480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_622.const", - "file_size": 1868800 - }, - "Attention_24_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1258137280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_623.const", - "file_size": 1868800 - }, - "Attention_24_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1260006080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_624.const", - "file_size": 1868800 - }, - "onnx::MatMul_17357_11_0_105": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1261874880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_625.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1263743680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_626.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1263746240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_627.const", - "file_size": 2560 - }, - "onnx::MatMul_17358": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1263748800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_628.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1265617600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_629.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1265617728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_630.const", - "file_size": 128 - }, - "onnx::MatMul_17368_11_0_106": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1265617856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_631.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1267486656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_632.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1267489216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_633.const", - "file_size": 2560 - }, - "onnx::MatMul_17369_11_0_107_12_24_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1267491776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_634.const", - "file_size": 7475200 - }, - "onnx::MatMul_17369_11_0_107_12_24_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1274966976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_635.const", - "file_size": 7475200 - }, - "onnx::MatMul_17370_11_0_108": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1282442176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_636.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1289917376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_637.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1289919936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_638.const", - "file_size": 2560 - }, - "Attention_25_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1289922496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_639.const", - "file_size": 1868800 - }, - "Attention_25_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1291791296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_640.const", - "file_size": 1868800 - }, - "Attention_25_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1293660096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_641.const", - "file_size": 1868800 - }, - "onnx::MatMul_17381_11_0_109": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1295528896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_642.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1297397696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_643.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1297400256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_644.const", - "file_size": 2560 - }, - "onnx::MatMul_17382": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1297402816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_645.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1299271616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_646.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1299271744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_647.const", - "file_size": 128 - }, - "onnx::MatMul_17392_11_0_110": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1299271872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_648.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1301140672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_649.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1301143232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_650.const", - "file_size": 2560 - }, - "onnx::MatMul_17393_11_0_111_12_25_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1301145792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_651.const", - "file_size": 7475200 - }, - "onnx::MatMul_17393_11_0_111_12_25_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1308620992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_652.const", - "file_size": 7475200 - }, - "onnx::MatMul_17394_11_0_112": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1316096192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_653.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1323571392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_654.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1323573952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_655.const", - "file_size": 2560 - }, - "Attention_26_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1323576512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_656.const", - "file_size": 1868800 - }, - "Attention_26_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1325445312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_657.const", - "file_size": 1868800 - }, - "Attention_26_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1327314112, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_658.const", - "file_size": 1868800 - }, - "onnx::MatMul_17405_11_0_113": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1329182912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_659.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1331051712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_660.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1331054272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_661.const", - "file_size": 2560 - }, - "onnx::MatMul_17406": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1331056832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_662.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1332925632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_663.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1332925760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_664.const", - "file_size": 128 - }, - "onnx::MatMul_17416_11_0_114": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1332925888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_665.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1334794688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_666.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1334797248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_667.const", - "file_size": 2560 - }, - "onnx::MatMul_17417_11_0_115_12_26_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1334799808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_668.const", - "file_size": 7475200 - }, - "onnx::MatMul_17417_11_0_115_12_26_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1342275008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_669.const", - "file_size": 7475200 - }, - "onnx::MatMul_17418_11_0_116": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1349750208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_670.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1357225408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_671.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1357227968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_672.const", - "file_size": 2560 - }, - "Attention_27_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1357230528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_673.const", - "file_size": 1868800 - }, - "Attention_27_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1359099328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_674.const", - "file_size": 1868800 - }, - "Attention_27_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1360968128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_675.const", - "file_size": 1868800 - }, - "onnx::MatMul_17429_11_0_117": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1362836928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_676.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1364705728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_677.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1364708288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_678.const", - "file_size": 2560 - }, - "onnx::MatMul_17430": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1364710848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_679.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1366579648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_680.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1366579776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_681.const", - "file_size": 128 - }, - "onnx::MatMul_17440_11_0_118": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1366579904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_682.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1368448704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_683.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1368451264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_684.const", - "file_size": 2560 - }, - "onnx::MatMul_17441_11_0_119_12_27_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1368453824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_685.const", - "file_size": 7475200 - }, - "onnx::MatMul_17441_11_0_119_12_27_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1375929024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_686.const", - "file_size": 7475200 - }, - "onnx::MatMul_17442_11_0_120": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1383404224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_687.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1390879424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_688.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1390881984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_689.const", - "file_size": 2560 - }, - "Attention_28_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1390884544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_690.const", - "file_size": 1868800 - }, - "Attention_28_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1392753344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_691.const", - "file_size": 1868800 - }, - "Attention_28_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1394622144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_692.const", - "file_size": 1868800 - }, - "onnx::MatMul_17453_11_0_121": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1396490944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_693.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1398359744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_694.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1398362304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_695.const", - "file_size": 2560 - }, - "onnx::MatMul_17454": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1398364864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_696.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1400233664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_697.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1400233792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_698.const", - "file_size": 128 - }, - "onnx::MatMul_17464_11_0_122": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1400233920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_699.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1402102720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_700.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1402105280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_701.const", - "file_size": 2560 - }, - "onnx::MatMul_17465_11_0_123_12_28_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1402107840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_702.const", - "file_size": 7475200 - }, - "onnx::MatMul_17465_11_0_123_12_28_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1409583040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_703.const", - "file_size": 7475200 - }, - "onnx::MatMul_17466_11_0_124": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1417058240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_704.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1424533440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_705.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1424536000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_706.const", - "file_size": 2560 - }, - "Attention_29_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1424538560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_707.const", - "file_size": 1868800 - }, - "Attention_29_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1426407360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_708.const", - "file_size": 1868800 - }, - "Attention_29_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1428276160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_709.const", - "file_size": 1868800 - }, - "onnx::MatMul_17477_11_0_125": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1430144960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_710.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1432013760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_711.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1432016320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_712.const", - "file_size": 2560 - }, - "onnx::MatMul_17478": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1432018880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_713.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1433887680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_714.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1433887808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_715.const", - "file_size": 128 - }, - "onnx::MatMul_17488_11_0_126": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1433887936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_716.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1435756736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_717.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1435759296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_718.const", - "file_size": 2560 - }, - "onnx::MatMul_17489_11_0_127_12_29_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1435761856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_719.const", - "file_size": 7475200 - }, - "onnx::MatMul_17489_11_0_127_12_29_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1443237056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_720.const", - "file_size": 7475200 - }, - "onnx::MatMul_17490_11_0_128": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1450712256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_721.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1458187456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_722.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1458190016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_723.const", - "file_size": 2560 - }, - "Attention_30_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1458192576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_724.const", - "file_size": 1868800 - }, - "Attention_30_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1460061376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_725.const", - "file_size": 1868800 - }, - "Attention_30_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1461930176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_726.const", - "file_size": 1868800 - }, - "onnx::MatMul_17501_11_0_129": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1463798976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_727.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1465667776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_728.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1465670336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_729.const", - "file_size": 2560 - }, - "onnx::MatMul_17502": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1465672896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_730.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1467541696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_731.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1467541824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_732.const", - "file_size": 128 - }, - "onnx::MatMul_17512_11_0_130": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1467541952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_733.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1469410752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_734.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1469413312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_735.const", - "file_size": 2560 - }, - "onnx::MatMul_17513_11_0_131_12_30_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1469415872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_736.const", - "file_size": 7475200 - }, - "onnx::MatMul_17513_11_0_131_12_30_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1476891072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_737.const", - "file_size": 7475200 - }, - "onnx::MatMul_17514_11_0_132": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1484366272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_738.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1491841472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_739.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1491844032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_740.const", - "file_size": 2560 - }, - "Attention_31_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1491846592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_741.const", - "file_size": 1868800 - }, - "Attention_31_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1493715392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_742.const", - "file_size": 1868800 - }, - "Attention_31_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1495584192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_743.const", - "file_size": 1868800 - }, - "onnx::MatMul_17525_11_0_133": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1497452992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_744.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1499321792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_745.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1499324352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_746.const", - "file_size": 2560 - }, - "onnx::MatMul_17526": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1499326912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_747.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1501195712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_748.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1501195840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_749.const", - "file_size": 128 - }, - "onnx::MatMul_17536_11_0_134": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1501195968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_750.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1503064768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_751.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1503067328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_752.const", - "file_size": 2560 - }, - "onnx::MatMul_17537_11_0_135_12_31_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1503069888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_753.const", - "file_size": 7475200 - }, - "onnx::MatMul_17537_11_0_135_12_31_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1510545088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_754.const", - "file_size": 7475200 - }, - "onnx::MatMul_17538_11_0_136": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1518020288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_755.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1525495488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_756.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1525498048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_757.const", - "file_size": 2560 - }, - "Attention_32_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1525500608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_758.const", - "file_size": 1868800 - }, - "Attention_32_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1527369408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_759.const", - "file_size": 1868800 - }, - "Attention_32_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1529238208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_760.const", - "file_size": 1868800 - }, - "onnx::MatMul_17549_11_0_137": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1531107008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_761.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1532975808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_762.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1532978368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_763.const", - "file_size": 2560 - }, - "onnx::MatMul_17550": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1532980928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_764.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1534849728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_765.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1534849856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_766.const", - "file_size": 128 - }, - "onnx::MatMul_17560_11_0_138": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1534849984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_767.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1536718784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_768.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1536721344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_769.const", - "file_size": 2560 - }, - "onnx::MatMul_17561_11_0_139_12_32_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1536723904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_770.const", - "file_size": 7475200 - }, - "onnx::MatMul_17561_11_0_139_12_32_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1544199104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_771.const", - "file_size": 7475200 - }, - "onnx::MatMul_17562_11_0_140": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1551674304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_772.const", - "file_size": 7475200 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1559149504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_773.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1559152064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_774.const", - "file_size": 2560 - }, - "Attention_33_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1559154624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_775.const", - "file_size": 1868800 - }, - "Attention_33_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1561023424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_776.const", - "file_size": 1868800 - }, - "Attention_33_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1562892224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_777.const", - "file_size": 1868800 - }, - "onnx::MatMul_17573_11_0_141": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1564761024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_778.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1566629824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_779.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1566632384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_780.const", - "file_size": 2560 - }, - "onnx::MatMul_17574": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1566634944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_781.const", - "file_size": 1868800 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1568503744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_782.const", - "file_size": 128 - }, - "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1568503872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_783.const", - "file_size": 128 - }, - "onnx::MatMul_17584_11_0_142": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1568504000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_784.const", - "file_size": 1868800 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1570372800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_785.const", - "file_size": 2560 - }, - "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1570375360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_786.const", - "file_size": 2560 - }, - "onnx::MatMul_17585_11_0_143_12_33_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1570377920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_787.const", - "file_size": 7475200 - }, - "onnx::MatMul_17585_11_0_143_12_33_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1577853120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_788.const", - "file_size": 7475200 - }, - "onnx::MatMul_17586_11_0_144": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1585328320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_789.const", - "file_size": 7475200 - }, - "onnx::MatMul_17587_11_0_145": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1592803520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_790.const", - "file_size": 1868800 - }, - "GroupNorm_19_wts_6_0_14": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1594672320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_791.const", - "file_size": 5120 - }, - "NhwcConv_19_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 1594677440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_792.const", - "file_size": 17244160 - }, - "GroupNorm_20_wts_6_0_15": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1611921600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_793.const", - "file_size": 5120 - }, - "NhwcConv_20_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 1611926720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_794.const", - "file_size": 17244160 - }, - "GroupNorm_21_wts_6_0_16": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 5120 - ], - "size_in_bytes": 10240, - "op_tensor_size": 10240, - "offset": 1629170880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_795.const", - "file_size": 10240 - }, - "NhwcConv_21_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 33832960 - ], - "size_in_bytes": 33832960, - "op_tensor_size": 33832960, - "offset": 1629181120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_796.const", - "file_size": 33832960 - }, - "GroupNorm_22_wts_6_0_17": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1663014080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_797.const", - "file_size": 5120 - }, - "NhwcConv_22_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 1663019200, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_798.const", - "file_size": 17244160 - }, - "NhwcConv_23_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4014080 - ], - "size_in_bytes": 4014080, - "op_tensor_size": 4014080, - "offset": 1680263360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_799.const", - "file_size": 4014080 - }, - "GroupNorm_23_wts_6_1_5": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 1684277440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_800.const", - "file_size": 5120 - }, - "onnx::MatMul_17598_11_0_146": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1684282560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_801.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1686151360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_802.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1686153920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_803.const", - "file_size": 2560 - }, - "Attention_34_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1686156480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_804.const", - "file_size": 1868800 - }, - "Attention_34_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1688025280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_805.const", - "file_size": 1868800 - }, - "Attention_34_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1689894080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_806.const", - "file_size": 1868800 - }, - "onnx::MatMul_17609_11_0_147": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1691762880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_807.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1693631680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_808.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1693634240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_809.const", - "file_size": 2560 - }, - "onnx::MatMul_17610": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1693636800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_810.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1695505600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_811.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1695505728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_812.const", - "file_size": 128 - }, - "onnx::MatMul_17620_11_0_148": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1695505856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_813.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1697374656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_814.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1697377216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_815.const", - "file_size": 2560 - }, - "onnx::MatMul_17621_11_0_149_12_34_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1697379776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_816.const", - "file_size": 7475200 - }, - "onnx::MatMul_17621_11_0_149_12_34_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1704854976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_817.const", - "file_size": 7475200 - }, - "onnx::MatMul_17622_11_0_150": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1712330176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_818.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1719805376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_819.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1719807936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_820.const", - "file_size": 2560 - }, - "Attention_35_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1719810496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_821.const", - "file_size": 1868800 - }, - "Attention_35_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1721679296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_822.const", - "file_size": 1868800 - }, - "Attention_35_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1723548096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_823.const", - "file_size": 1868800 - }, - "onnx::MatMul_17633_11_0_151": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1725416896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_824.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1727285696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_825.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1727288256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_826.const", - "file_size": 2560 - }, - "onnx::MatMul_17634": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1727290816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_827.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1729159616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_828.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1729159744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_829.const", - "file_size": 128 - }, - "onnx::MatMul_17644_11_0_152": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1729159872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_830.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1731028672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_831.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1731031232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_832.const", - "file_size": 2560 - }, - "onnx::MatMul_17645_11_0_153_12_35_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1731033792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_833.const", - "file_size": 7475200 - }, - "onnx::MatMul_17645_11_0_153_12_35_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1738508992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_834.const", - "file_size": 7475200 - }, - "onnx::MatMul_17646_11_0_154": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1745984192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_835.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1753459392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_836.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1753461952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_837.const", - "file_size": 2560 - }, - "Attention_36_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1753464512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_838.const", - "file_size": 1868800 - }, - "Attention_36_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1755333312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_839.const", - "file_size": 1868800 - }, - "Attention_36_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1757202112, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_840.const", - "file_size": 1868800 - }, - "onnx::MatMul_17657_11_0_155": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1759070912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_841.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1760939712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_842.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1760942272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_843.const", - "file_size": 2560 - }, - "onnx::MatMul_17658": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1760944832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_844.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1762813632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_845.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1762813760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_846.const", - "file_size": 128 - }, - "onnx::MatMul_17668_11_0_156": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1762813888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_847.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1764682688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_848.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1764685248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_849.const", - "file_size": 2560 - }, - "onnx::MatMul_17669_11_0_157_12_36_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1764687808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_850.const", - "file_size": 7475200 - }, - "onnx::MatMul_17669_11_0_157_12_36_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1772163008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_851.const", - "file_size": 7475200 - }, - "onnx::MatMul_17670_11_0_158": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1779638208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_852.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1787113408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_853.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1787115968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_854.const", - "file_size": 2560 - }, - "Attention_37_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1787118528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_855.const", - "file_size": 1868800 - }, - "Attention_37_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1788987328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_856.const", - "file_size": 1868800 - }, - "Attention_37_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1790856128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_857.const", - "file_size": 1868800 - }, - "onnx::MatMul_17681_11_0_159": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1792724928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_858.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1794593728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_859.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1794596288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_860.const", - "file_size": 2560 - }, - "onnx::MatMul_17682": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1794598848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_861.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1796467648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_862.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1796467776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_863.const", - "file_size": 128 - }, - "onnx::MatMul_17692_11_0_160": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1796467904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_864.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1798336704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_865.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1798339264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_866.const", - "file_size": 2560 - }, - "onnx::MatMul_17693_11_0_161_12_37_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1798341824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_867.const", - "file_size": 7475200 - }, - "onnx::MatMul_17693_11_0_161_12_37_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1805817024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_868.const", - "file_size": 7475200 - }, - "onnx::MatMul_17694_11_0_162": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1813292224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_869.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1820767424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_870.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1820769984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_871.const", - "file_size": 2560 - }, - "Attention_38_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1820772544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_872.const", - "file_size": 1868800 - }, - "Attention_38_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1822641344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_873.const", - "file_size": 1868800 - }, - "Attention_38_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1824510144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_874.const", - "file_size": 1868800 - }, - "onnx::MatMul_17705_11_0_163": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1826378944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_875.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1828247744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_876.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1828250304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_877.const", - "file_size": 2560 - }, - "onnx::MatMul_17706": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1828252864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_878.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1830121664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_879.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1830121792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_880.const", - "file_size": 128 - }, - "onnx::MatMul_17716_11_0_164": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1830121920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_881.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1831990720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_882.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1831993280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_883.const", - "file_size": 2560 - }, - "onnx::MatMul_17717_11_0_165_12_38_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1831995840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_884.const", - "file_size": 7475200 - }, - "onnx::MatMul_17717_11_0_165_12_38_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1839471040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_885.const", - "file_size": 7475200 - }, - "onnx::MatMul_17718_11_0_166": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1846946240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_886.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1854421440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_887.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1854424000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_888.const", - "file_size": 2560 - }, - "Attention_39_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1854426560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_889.const", - "file_size": 1868800 - }, - "Attention_39_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1856295360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_890.const", - "file_size": 1868800 - }, - "Attention_39_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1858164160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_891.const", - "file_size": 1868800 - }, - "onnx::MatMul_17729_11_0_167": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1860032960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_892.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1861901760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_893.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1861904320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_894.const", - "file_size": 2560 - }, - "onnx::MatMul_17730": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1861906880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_895.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1863775680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_896.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1863775808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_897.const", - "file_size": 128 - }, - "onnx::MatMul_17740_11_0_168": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1863775936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_898.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1865644736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_899.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1865647296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_900.const", - "file_size": 2560 - }, - "onnx::MatMul_17741_11_0_169_12_39_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1865649856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_901.const", - "file_size": 7475200 - }, - "onnx::MatMul_17741_11_0_169_12_39_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1873125056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_902.const", - "file_size": 7475200 - }, - "onnx::MatMul_17742_11_0_170": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1880600256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_903.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1888075456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_904.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1888078016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_905.const", - "file_size": 2560 - }, - "Attention_40_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1888080576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_906.const", - "file_size": 1868800 - }, - "Attention_40_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1889949376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_907.const", - "file_size": 1868800 - }, - "Attention_40_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1891818176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_908.const", - "file_size": 1868800 - }, - "onnx::MatMul_17753_11_0_171": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1893686976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_909.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1895555776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_910.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1895558336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_911.const", - "file_size": 2560 - }, - "onnx::MatMul_17754": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1895560896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_912.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1897429696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_913.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1897429824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_914.const", - "file_size": 128 - }, - "onnx::MatMul_17764_11_0_172": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1897429952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_915.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1899298752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_916.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1899301312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_917.const", - "file_size": 2560 - }, - "onnx::MatMul_17765_11_0_173_12_40_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1899303872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_918.const", - "file_size": 7475200 - }, - "onnx::MatMul_17765_11_0_173_12_40_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1906779072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_919.const", - "file_size": 7475200 - }, - "onnx::MatMul_17766_11_0_174": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1914254272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_920.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1921729472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_921.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1921732032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_922.const", - "file_size": 2560 - }, - "Attention_41_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1921734592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_923.const", - "file_size": 1868800 - }, - "Attention_41_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1923603392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_924.const", - "file_size": 1868800 - }, - "Attention_41_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1925472192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_925.const", - "file_size": 1868800 - }, - "onnx::MatMul_17777_11_0_175": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1927340992, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_926.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1929209792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_927.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1929212352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_928.const", - "file_size": 2560 - }, - "onnx::MatMul_17778": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1929214912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_929.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1931083712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_930.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1931083840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_931.const", - "file_size": 128 - }, - "onnx::MatMul_17788_11_0_176": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1931083968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_932.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1932952768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_933.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1932955328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_934.const", - "file_size": 2560 - }, - "onnx::MatMul_17789_11_0_177_12_41_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1932957888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_935.const", - "file_size": 7475200 - }, - "onnx::MatMul_17789_11_0_177_12_41_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1940433088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_936.const", - "file_size": 7475200 - }, - "onnx::MatMul_17790_11_0_178": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1947908288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_937.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1955383488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_938.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1955386048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_939.const", - "file_size": 2560 - }, - "Attention_42_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1955388608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_940.const", - "file_size": 1868800 - }, - "Attention_42_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1957257408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_941.const", - "file_size": 1868800 - }, - "Attention_42_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1959126208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_942.const", - "file_size": 1868800 - }, - "onnx::MatMul_17801_11_0_179": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1960995008, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_943.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1962863808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_944.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1962866368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_945.const", - "file_size": 2560 - }, - "onnx::MatMul_17802": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1962868928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_946.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1964737728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_947.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1964737856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_948.const", - "file_size": 128 - }, - "onnx::MatMul_17812_11_0_180": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1964737984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_949.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1966606784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_950.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1966609344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_951.const", - "file_size": 2560 - }, - "onnx::MatMul_17813_11_0_181_12_42_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1966611904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_952.const", - "file_size": 7475200 - }, - "onnx::MatMul_17813_11_0_181_12_42_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1974087104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_953.const", - "file_size": 7475200 - }, - "onnx::MatMul_17814_11_0_182": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 1981562304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_954.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1989037504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_955.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1989040064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_956.const", - "file_size": 2560 - }, - "Attention_43_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1989042624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_957.const", - "file_size": 1868800 - }, - "Attention_43_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1990911424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_958.const", - "file_size": 1868800 - }, - "Attention_43_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1992780224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_959.const", - "file_size": 1868800 - }, - "onnx::MatMul_17825_11_0_183": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1994649024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_960.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1996517824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_961.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 1996520384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_962.const", - "file_size": 2560 - }, - "onnx::MatMul_17826": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1996522944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_963.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1998391744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_964.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 1998391872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_965.const", - "file_size": 128 - }, - "onnx::MatMul_17836_11_0_184": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 1998392000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_966.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2000260800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_967.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2000263360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_968.const", - "file_size": 2560 - }, - "onnx::MatMul_17837_11_0_185_12_43_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2000265920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_969.const", - "file_size": 7475200 - }, - "onnx::MatMul_17837_11_0_185_12_43_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2007741120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_970.const", - "file_size": 7475200 - }, - "onnx::MatMul_17838_11_0_186": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2015216320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_971.const", - "file_size": 7475200 - }, - "onnx::MatMul_17839_11_0_187": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2022691520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_972.const", - "file_size": 1868800 - }, - "GroupNorm_24_wts_6_0_18": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 5120 - ], - "size_in_bytes": 10240, - "op_tensor_size": 10240, - "offset": 2024560320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_973.const", - "file_size": 10240 - }, - "NhwcConv_24_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 33832960 - ], - "size_in_bytes": 33832960, - "op_tensor_size": 33832960, - "offset": 2024570560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_974.const", - "file_size": 33832960 - }, - "GroupNorm_25_wts_6_0_19": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 2058403520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_975.const", - "file_size": 5120 - }, - "NhwcConv_25_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 2058408640, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_976.const", - "file_size": 17244160 - }, - "NhwcConv_26_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4014080 - ], - "size_in_bytes": 4014080, - "op_tensor_size": 4014080, - "offset": 2075652800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_977.const", - "file_size": 4014080 - }, - "GroupNorm_26_wts_6_1_6": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 2079666880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_978.const", - "file_size": 5120 - }, - "onnx::MatMul_17846_11_0_188": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2079672000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_979.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2081540800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_980.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2081543360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_981.const", - "file_size": 2560 - }, - "Attention_44_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2081545920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_982.const", - "file_size": 1868800 - }, - "Attention_44_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2083414720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_983.const", - "file_size": 1868800 - }, - "Attention_44_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2085283520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_984.const", - "file_size": 1868800 - }, - "onnx::MatMul_17857_11_0_189": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2087152320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_985.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2089021120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_986.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2089023680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_987.const", - "file_size": 2560 - }, - "onnx::MatMul_17858": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2089026240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_988.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2090895040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_989.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2090895168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_990.const", - "file_size": 128 - }, - "onnx::MatMul_17868_11_0_190": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2090895296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_991.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2092764096, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_992.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2092766656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_993.const", - "file_size": 2560 - }, - "onnx::MatMul_17869_11_0_191_12_44_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2092769216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_994.const", - "file_size": 7475200 - }, - "onnx::MatMul_17869_11_0_191_12_44_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2100244416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_995.const", - "file_size": 7475200 - }, - "onnx::MatMul_17870_11_0_192": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2107719616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_996.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2115194816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_997.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2115197376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_998.const", - "file_size": 2560 - }, - "Attention_45_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2115199936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_999.const", - "file_size": 1868800 - }, - "Attention_45_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2117068736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1000.const", - "file_size": 1868800 - }, - "Attention_45_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2118937536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1001.const", - "file_size": 1868800 - }, - "onnx::MatMul_17881_11_0_193": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2120806336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1002.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2122675136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1003.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2122677696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1004.const", - "file_size": 2560 - }, - "onnx::MatMul_17882": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2122680256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1005.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2124549056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1006.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2124549184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1007.const", - "file_size": 128 - }, - "onnx::MatMul_17892_11_0_194": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2124549312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1008.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2126418112, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1009.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2126420672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1010.const", - "file_size": 2560 - }, - "onnx::MatMul_17893_11_0_195_12_45_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2126423232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1011.const", - "file_size": 7475200 - }, - "onnx::MatMul_17893_11_0_195_12_45_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2133898432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1012.const", - "file_size": 7475200 - }, - "onnx::MatMul_17894_11_0_196": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2141373632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1013.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2148848832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1014.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2148851392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1015.const", - "file_size": 2560 - }, - "Attention_46_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2148853952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1016.const", - "file_size": 1868800 - }, - "Attention_46_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2150722752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1017.const", - "file_size": 1868800 - }, - "Attention_46_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2152591552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1018.const", - "file_size": 1868800 - }, - "onnx::MatMul_17905_11_0_197": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2154460352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1019.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2156329152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1020.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2156331712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1021.const", - "file_size": 2560 - }, - "onnx::MatMul_17906": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2156334272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1022.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2158203072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1023.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2158203200, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1024.const", - "file_size": 128 - }, - "onnx::MatMul_17916_11_0_198": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2158203328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1025.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2160072128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1026.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2160074688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1027.const", - "file_size": 2560 - }, - "onnx::MatMul_17917_11_0_199_12_46_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2160077248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1028.const", - "file_size": 7475200 - }, - "onnx::MatMul_17917_11_0_199_12_46_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2167552448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1029.const", - "file_size": 7475200 - }, - "onnx::MatMul_17918_11_0_200": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2175027648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1030.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2182502848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1031.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2182505408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1032.const", - "file_size": 2560 - }, - "Attention_47_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2182507968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1033.const", - "file_size": 1868800 - }, - "Attention_47_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2184376768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1034.const", - "file_size": 1868800 - }, - "Attention_47_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2186245568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1035.const", - "file_size": 1868800 - }, - "onnx::MatMul_17929_11_0_201": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2188114368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1036.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2189983168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1037.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2189985728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1038.const", - "file_size": 2560 - }, - "onnx::MatMul_17930": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2189988288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1039.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2191857088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1040.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2191857216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1041.const", - "file_size": 128 - }, - "onnx::MatMul_17940_11_0_202": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2191857344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1042.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2193726144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1043.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2193728704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1044.const", - "file_size": 2560 - }, - "onnx::MatMul_17941_11_0_203_12_47_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2193731264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1045.const", - "file_size": 7475200 - }, - "onnx::MatMul_17941_11_0_203_12_47_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2201206464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1046.const", - "file_size": 7475200 - }, - "onnx::MatMul_17942_11_0_204": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2208681664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1047.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2216156864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1048.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2216159424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1049.const", - "file_size": 2560 - }, - "Attention_48_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2216161984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1050.const", - "file_size": 1868800 - }, - "Attention_48_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2218030784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1051.const", - "file_size": 1868800 - }, - "Attention_48_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2219899584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1052.const", - "file_size": 1868800 - }, - "onnx::MatMul_17953_11_0_205": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2221768384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1053.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2223637184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1054.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2223639744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1055.const", - "file_size": 2560 - }, - "onnx::MatMul_17954": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2223642304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1056.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2225511104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1057.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2225511232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1058.const", - "file_size": 128 - }, - "onnx::MatMul_17964_11_0_206": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2225511360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1059.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2227380160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1060.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2227382720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1061.const", - "file_size": 2560 - }, - "onnx::MatMul_17965_11_0_207_12_48_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2227385280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1062.const", - "file_size": 7475200 - }, - "onnx::MatMul_17965_11_0_207_12_48_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2234860480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1063.const", - "file_size": 7475200 - }, - "onnx::MatMul_17966_11_0_208": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2242335680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1064.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2249810880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1065.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2249813440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1066.const", - "file_size": 2560 - }, - "Attention_49_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2249816000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1067.const", - "file_size": 1868800 - }, - "Attention_49_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2251684800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1068.const", - "file_size": 1868800 - }, - "Attention_49_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2253553600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1069.const", - "file_size": 1868800 - }, - "onnx::MatMul_17977_11_0_209": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2255422400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1070.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2257291200, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1071.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2257293760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1072.const", - "file_size": 2560 - }, - "onnx::MatMul_17978": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2257296320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1073.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2259165120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1074.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2259165248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1075.const", - "file_size": 128 - }, - "onnx::MatMul_17988_11_0_210": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2259165376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1076.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2261034176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1077.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2261036736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1078.const", - "file_size": 2560 - }, - "onnx::MatMul_17989_11_0_211_12_49_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2261039296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1079.const", - "file_size": 7475200 - }, - "onnx::MatMul_17989_11_0_211_12_49_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2268514496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1080.const", - "file_size": 7475200 - }, - "onnx::MatMul_17990_11_0_212": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2275989696, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1081.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2283464896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1082.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2283467456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1083.const", - "file_size": 2560 - }, - "Attention_50_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2283470016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1084.const", - "file_size": 1868800 - }, - "Attention_50_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2285338816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1085.const", - "file_size": 1868800 - }, - "Attention_50_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2287207616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1086.const", - "file_size": 1868800 - }, - "onnx::MatMul_18001_11_0_213": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2289076416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1087.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2290945216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1088.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2290947776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1089.const", - "file_size": 2560 - }, - "onnx::MatMul_18002": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2290950336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1090.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2292819136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1091.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2292819264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1092.const", - "file_size": 128 - }, - "onnx::MatMul_18012_11_0_214": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2292819392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1093.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2294688192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1094.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2294690752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1095.const", - "file_size": 2560 - }, - "onnx::MatMul_18013_11_0_215_12_50_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2294693312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1096.const", - "file_size": 7475200 - }, - "onnx::MatMul_18013_11_0_215_12_50_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2302168512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1097.const", - "file_size": 7475200 - }, - "onnx::MatMul_18014_11_0_216": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2309643712, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1098.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2317118912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1099.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2317121472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1100.const", - "file_size": 2560 - }, - "Attention_51_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2317124032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1101.const", - "file_size": 1868800 - }, - "Attention_51_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2318992832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1102.const", - "file_size": 1868800 - }, - "Attention_51_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2320861632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1103.const", - "file_size": 1868800 - }, - "onnx::MatMul_18025_11_0_217": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2322730432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1104.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2324599232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1105.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2324601792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1106.const", - "file_size": 2560 - }, - "onnx::MatMul_18026": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2324604352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1107.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2326473152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1108.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2326473280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1109.const", - "file_size": 128 - }, - "onnx::MatMul_18036_11_0_218": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2326473408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1110.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2328342208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1111.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2328344768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1112.const", - "file_size": 2560 - }, - "onnx::MatMul_18037_11_0_219_12_51_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2328347328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1113.const", - "file_size": 7475200 - }, - "onnx::MatMul_18037_11_0_219_12_51_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2335822528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1114.const", - "file_size": 7475200 - }, - "onnx::MatMul_18038_11_0_220": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2343297728, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1115.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2350772928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1116.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2350775488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1117.const", - "file_size": 2560 - }, - "Attention_52_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2350778048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1118.const", - "file_size": 1868800 - }, - "Attention_52_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2352646848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1119.const", - "file_size": 1868800 - }, - "Attention_52_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2354515648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1120.const", - "file_size": 1868800 - }, - "onnx::MatMul_18049_11_0_221": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2356384448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1121.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2358253248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1122.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2358255808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1123.const", - "file_size": 2560 - }, - "onnx::MatMul_18050": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2358258368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1124.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2360127168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1125.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2360127296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1126.const", - "file_size": 128 - }, - "onnx::MatMul_18060_11_0_222": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2360127424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1127.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2361996224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1128.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2361998784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1129.const", - "file_size": 2560 - }, - "onnx::MatMul_18061_11_0_223_12_52_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2362001344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1130.const", - "file_size": 7475200 - }, - "onnx::MatMul_18061_11_0_223_12_52_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2369476544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1131.const", - "file_size": 7475200 - }, - "onnx::MatMul_18062_11_0_224": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2376951744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1132.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2384426944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1133.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2384429504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1134.const", - "file_size": 2560 - }, - "Attention_53_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2384432064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1135.const", - "file_size": 1868800 - }, - "Attention_53_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2386300864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1136.const", - "file_size": 1868800 - }, - "Attention_53_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2388169664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1137.const", - "file_size": 1868800 - }, - "onnx::MatMul_18073_11_0_225": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2390038464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1138.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2391907264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1139.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2391909824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1140.const", - "file_size": 2560 - }, - "onnx::MatMul_18074": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2391912384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1141.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2393781184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1142.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2393781312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1143.const", - "file_size": 128 - }, - "onnx::MatMul_18084_11_0_226": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2393781440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1144.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2395650240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1145.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2395652800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1146.const", - "file_size": 2560 - }, - "onnx::MatMul_18085_11_0_227_12_53_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2395655360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1147.const", - "file_size": 7475200 - }, - "onnx::MatMul_18085_11_0_227_12_53_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2403130560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1148.const", - "file_size": 7475200 - }, - "onnx::MatMul_18086_11_0_228": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2410605760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1149.const", - "file_size": 7475200 - }, - "onnx::MatMul_18087_11_0_229": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2418080960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1150.const", - "file_size": 1868800 - }, - "GroupNorm_27_wts_6_0_20": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 3840 - ], - "size_in_bytes": 7680, - "op_tensor_size": 7680, - "offset": 2419949760, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1151.const", - "file_size": 7680 - }, - "NhwcConv_27_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 25374720 - ], - "size_in_bytes": 25374720, - "op_tensor_size": 25374720, - "offset": 2419957440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1152.const", - "file_size": 25374720 - }, - "GroupNorm_28_wts_6_0_21": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 2445332160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1153.const", - "file_size": 5120 - }, - "NhwcConv_28_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 2445337280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1154.const", - "file_size": 17244160 - }, - "NhwcConv_29_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 3010560 - ], - "size_in_bytes": 3010560, - "op_tensor_size": 3010560, - "offset": 2462581440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1155.const", - "file_size": 3010560 - }, - "GroupNorm_29_wts_6_1_7": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 2465592000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1156.const", - "file_size": 5120 - }, - "onnx::MatMul_18094_11_0_230": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2465597120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1157.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2467465920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1158.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2467468480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1159.const", - "file_size": 2560 - }, - "Attention_54_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2467471040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1160.const", - "file_size": 1868800 - }, - "Attention_54_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2469339840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1161.const", - "file_size": 1868800 - }, - "Attention_54_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2471208640, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1162.const", - "file_size": 1868800 - }, - "onnx::MatMul_18105_11_0_231": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2473077440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1163.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2474946240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1164.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2474948800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1165.const", - "file_size": 2560 - }, - "onnx::MatMul_18106": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2474951360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1166.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2476820160, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1167.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2476820288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1168.const", - "file_size": 128 - }, - "onnx::MatMul_18116_11_0_232": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2476820416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1169.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2478689216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1170.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2478691776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1171.const", - "file_size": 2560 - }, - "onnx::MatMul_18117_11_0_233_12_54_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2478694336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1172.const", - "file_size": 7475200 - }, - "onnx::MatMul_18117_11_0_233_12_54_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2486169536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1173.const", - "file_size": 7475200 - }, - "onnx::MatMul_18118_11_0_234": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2493644736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1174.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2501119936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1175.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2501122496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1176.const", - "file_size": 2560 - }, - "Attention_55_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2501125056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1177.const", - "file_size": 1868800 - }, - "Attention_55_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2502993856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1178.const", - "file_size": 1868800 - }, - "Attention_55_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2504862656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1179.const", - "file_size": 1868800 - }, - "onnx::MatMul_18129_11_0_235": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2506731456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1180.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2508600256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1181.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2508602816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1182.const", - "file_size": 2560 - }, - "onnx::MatMul_18130": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2508605376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1183.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2510474176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1184.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2510474304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1185.const", - "file_size": 128 - }, - "onnx::MatMul_18140_11_0_236": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2510474432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1186.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2512343232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1187.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2512345792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1188.const", - "file_size": 2560 - }, - "onnx::MatMul_18141_11_0_237_12_55_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2512348352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1189.const", - "file_size": 7475200 - }, - "onnx::MatMul_18141_11_0_237_12_55_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2519823552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1190.const", - "file_size": 7475200 - }, - "onnx::MatMul_18142_11_0_238": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2527298752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1191.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2534773952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1192.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2534776512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1193.const", - "file_size": 2560 - }, - "Attention_56_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2534779072, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1194.const", - "file_size": 1868800 - }, - "Attention_56_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2536647872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1195.const", - "file_size": 1868800 - }, - "Attention_56_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2538516672, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1196.const", - "file_size": 1868800 - }, - "onnx::MatMul_18153_11_0_239": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2540385472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1197.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2542254272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1198.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2542256832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1199.const", - "file_size": 2560 - }, - "onnx::MatMul_18154": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2542259392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1200.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2544128192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1201.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2544128320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1202.const", - "file_size": 128 - }, - "onnx::MatMul_18164_11_0_240": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2544128448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1203.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2545997248, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1204.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2545999808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1205.const", - "file_size": 2560 - }, - "onnx::MatMul_18165_11_0_241_12_56_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2546002368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1206.const", - "file_size": 7475200 - }, - "onnx::MatMul_18165_11_0_241_12_56_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2553477568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1207.const", - "file_size": 7475200 - }, - "onnx::MatMul_18166_11_0_242": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2560952768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1208.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2568427968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1209.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2568430528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1210.const", - "file_size": 2560 - }, - "Attention_57_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2568433088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1211.const", - "file_size": 1868800 - }, - "Attention_57_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2570301888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1212.const", - "file_size": 1868800 - }, - "Attention_57_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2572170688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1213.const", - "file_size": 1868800 - }, - "onnx::MatMul_18177_11_0_243": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2574039488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1214.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2575908288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1215.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2575910848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1216.const", - "file_size": 2560 - }, - "onnx::MatMul_18178": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2575913408, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1217.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2577782208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1218.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2577782336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1219.const", - "file_size": 128 - }, - "onnx::MatMul_18188_11_0_244": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2577782464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1220.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2579651264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1221.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2579653824, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1222.const", - "file_size": 2560 - }, - "onnx::MatMul_18189_11_0_245_12_57_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2579656384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1223.const", - "file_size": 7475200 - }, - "onnx::MatMul_18189_11_0_245_12_57_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2587131584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1224.const", - "file_size": 7475200 - }, - "onnx::MatMul_18190_11_0_246": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2594606784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1225.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2602081984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1226.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2602084544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1227.const", - "file_size": 2560 - }, - "Attention_58_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2602087104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1228.const", - "file_size": 1868800 - }, - "Attention_58_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2603955904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1229.const", - "file_size": 1868800 - }, - "Attention_58_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2605824704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1230.const", - "file_size": 1868800 - }, - "onnx::MatMul_18201_11_0_247": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2607693504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1231.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2609562304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1232.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2609564864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1233.const", - "file_size": 2560 - }, - "onnx::MatMul_18202": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2609567424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1234.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2611436224, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1235.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2611436352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1236.const", - "file_size": 128 - }, - "onnx::MatMul_18212_11_0_248": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2611436480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1237.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2613305280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1238.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2613307840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1239.const", - "file_size": 2560 - }, - "onnx::MatMul_18213_11_0_249_12_58_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2613310400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1240.const", - "file_size": 7475200 - }, - "onnx::MatMul_18213_11_0_249_12_58_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2620785600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1241.const", - "file_size": 7475200 - }, - "onnx::MatMul_18214_11_0_250": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2628260800, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1242.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2635736000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1243.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2635738560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1244.const", - "file_size": 2560 - }, - "Attention_59_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2635741120, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1245.const", - "file_size": 1868800 - }, - "Attention_59_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2637609920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1246.const", - "file_size": 1868800 - }, - "Attention_59_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2639478720, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1247.const", - "file_size": 1868800 - }, - "onnx::MatMul_18225_11_0_251": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2641347520, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1248.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2643216320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1249.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2643218880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1250.const", - "file_size": 2560 - }, - "onnx::MatMul_18226": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2643221440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1251.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2645090240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1252.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2645090368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1253.const", - "file_size": 128 - }, - "onnx::MatMul_18236_11_0_252": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2645090496, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1254.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2646959296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1255.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2646961856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1256.const", - "file_size": 2560 - }, - "onnx::MatMul_18237_11_0_253_12_59_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2646964416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1257.const", - "file_size": 7475200 - }, - "onnx::MatMul_18237_11_0_253_12_59_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2654439616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1258.const", - "file_size": 7475200 - }, - "onnx::MatMul_18238_11_0_254": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2661914816, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1259.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2669390016, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1260.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2669392576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1261.const", - "file_size": 2560 - }, - "Attention_60_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2669395136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1262.const", - "file_size": 1868800 - }, - "Attention_60_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2671263936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1263.const", - "file_size": 1868800 - }, - "Attention_60_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2673132736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1264.const", - "file_size": 1868800 - }, - "onnx::MatMul_18249_11_0_255": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2675001536, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1265.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2676870336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1266.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2676872896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1267.const", - "file_size": 2560 - }, - "onnx::MatMul_18250": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2676875456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1268.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2678744256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1269.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2678744384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1270.const", - "file_size": 128 - }, - "onnx::MatMul_18260_11_0_256": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2678744512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1271.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2680613312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1272.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2680615872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1273.const", - "file_size": 2560 - }, - "onnx::MatMul_18261_11_0_257_12_60_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2680618432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1274.const", - "file_size": 7475200 - }, - "onnx::MatMul_18261_11_0_257_12_60_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2688093632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1275.const", - "file_size": 7475200 - }, - "onnx::MatMul_18262_11_0_258": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2695568832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1276.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2703044032, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1277.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2703046592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1278.const", - "file_size": 2560 - }, - "Attention_61_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2703049152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1279.const", - "file_size": 1868800 - }, - "Attention_61_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2704917952, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1280.const", - "file_size": 1868800 - }, - "Attention_61_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2706786752, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1281.const", - "file_size": 1868800 - }, - "onnx::MatMul_18273_11_0_259": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2708655552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1282.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2710524352, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1283.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2710526912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1284.const", - "file_size": 2560 - }, - "onnx::MatMul_18274": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2710529472, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1285.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2712398272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1286.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2712398400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1287.const", - "file_size": 128 - }, - "onnx::MatMul_18284_11_0_260": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2712398528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1288.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2714267328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1289.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2714269888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1290.const", - "file_size": 2560 - }, - "onnx::MatMul_18285_11_0_261_12_61_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2714272448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1291.const", - "file_size": 7475200 - }, - "onnx::MatMul_18285_11_0_261_12_61_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2721747648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1292.const", - "file_size": 7475200 - }, - "onnx::MatMul_18286_11_0_262": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2729222848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1293.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2736698048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1294.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2736700608, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1295.const", - "file_size": 2560 - }, - "Attention_62_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2736703168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1296.const", - "file_size": 1868800 - }, - "Attention_62_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2738571968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1297.const", - "file_size": 1868800 - }, - "Attention_62_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2740440768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1298.const", - "file_size": 1868800 - }, - "onnx::MatMul_18297_11_0_263": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2742309568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1299.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2744178368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1300.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2744180928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1301.const", - "file_size": 2560 - }, - "onnx::MatMul_18298": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2744183488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1302.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2746052288, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1303.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2746052416, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1304.const", - "file_size": 128 - }, - "onnx::MatMul_18308_11_0_264": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2746052544, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1305.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2747921344, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1306.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2747923904, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1307.const", - "file_size": 2560 - }, - "onnx::MatMul_18309_11_0_265_12_62_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2747926464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1308.const", - "file_size": 7475200 - }, - "onnx::MatMul_18309_11_0_265_12_62_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2755401664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1309.const", - "file_size": 7475200 - }, - "onnx::MatMul_18310_11_0_266": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2762876864, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1310.const", - "file_size": 7475200 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2770352064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1311.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2770354624, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1312.const", - "file_size": 2560 - }, - "Attention_63_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2770357184, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1313.const", - "file_size": 1868800 - }, - "Attention_63_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2772225984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1314.const", - "file_size": 1868800 - }, - "Attention_63_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2774094784, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1315.const", - "file_size": 1868800 - }, - "onnx::MatMul_18321_11_0_267": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2775963584, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1316.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2777832384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1317.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2777834944, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1318.const", - "file_size": 2560 - }, - "onnx::MatMul_18322": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2777837504, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1319.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2779706304, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1320.const", - "file_size": 128 - }, - "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2779706432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1321.const", - "file_size": 128 - }, - "onnx::MatMul_18332_11_0_268": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2779706560, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1322.const", - "file_size": 1868800 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2781575360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1323.const", - "file_size": 2560 - }, - "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2781577920, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1324.const", - "file_size": 2560 - }, - "onnx::MatMul_18333_11_0_269_12_63_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2781580480, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1325.const", - "file_size": 7475200 - }, - "onnx::MatMul_18333_11_0_269_12_63_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2789055680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1326.const", - "file_size": 7475200 - }, - "onnx::MatMul_18334_11_0_270": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 7475200 - ], - "size_in_bytes": 7475200, - "op_tensor_size": 7475200, - "offset": 2796530880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1327.const", - "file_size": 7475200 - }, - "onnx::MatMul_18335_11_0_271": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2804006080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1328.const", - "file_size": 1868800 - }, - "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 128 - ], - "size_in_bytes": 256, - "op_tensor_size": 256, - "offset": 2805874880, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1329.const", - "file_size": 256 - }, - "NhwcConv_30_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 17244160 - ], - "size_in_bytes": 17244160, - "op_tensor_size": 17244160, - "offset": 2805875136, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1330.const", - "file_size": 17244160 - }, - "GroupNorm_30_wts_6_0_22": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 3840 - ], - "size_in_bytes": 7680, - "op_tensor_size": 7680, - "offset": 2823119296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1331.const", - "file_size": 7680 - }, - "NhwcConv_31_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 12851200 - ], - "size_in_bytes": 12851200, - "op_tensor_size": 12851200, - "offset": 2823126976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1332.const", - "file_size": 12851200 - }, - "GroupNorm_31_wts_6_0_23": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2835978176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1333.const", - "file_size": 2560 - }, - "NhwcConv_32_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 2835980736, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1334.const", - "file_size": 4311040 - }, - "NhwcConv_33_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1505280 - ], - "size_in_bytes": 1505280, - "op_tensor_size": 1505280, - "offset": 2840291776, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1335.const", - "file_size": 1505280 - }, - "GroupNorm_32_wts_6_1_8": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2841797056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1336.const", - "file_size": 2560 - }, - "onnx::MatMul_18343_11_0_272": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2841799616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1337.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2842270656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1338.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2842271936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1339.const", - "file_size": 1280 - }, - "Attention_64_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2842273216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1340.const", - "file_size": 471040 - }, - "Attention_64_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2842744256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1341.const", - "file_size": 471040 - }, - "Attention_64_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2843215296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1342.const", - "file_size": 471040 - }, - "onnx::MatMul_18354_11_0_273": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2843686336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1343.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2844157376, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1344.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2844158656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1345.const", - "file_size": 1280 - }, - "onnx::MatMul_18355": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2844159936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1346.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2844630976, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1347.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2844631104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1348.const", - "file_size": 128 - }, - "onnx::MatMul_18365_11_0_274": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2844631232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1349.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2845102272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1350.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2845103552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1351.const", - "file_size": 1280 - }, - "onnx::MatMul_18366_11_0_275_12_64_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2845104832, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1352.const", - "file_size": 1868800 - }, - "onnx::MatMul_18366_11_0_275_12_64_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2846973632, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1353.const", - "file_size": 1868800 - }, - "onnx::MatMul_18367_11_0_276": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2848842432, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1354.const", - "file_size": 1884160 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2850726592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1355.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2850727872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1356.const", - "file_size": 1280 - }, - "Attention_65_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2850729152, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1357.const", - "file_size": 471040 - }, - "Attention_65_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2851200192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1358.const", - "file_size": 471040 - }, - "Attention_65_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2851671232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1359.const", - "file_size": 471040 - }, - "onnx::MatMul_18378_11_0_277": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2852142272, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1360.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2852613312, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1361.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2852614592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1362.const", - "file_size": 1280 - }, - "onnx::MatMul_18379": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2852615872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1363.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2853086912, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1364.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2853087040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1365.const", - "file_size": 128 - }, - "onnx::MatMul_18389_11_0_278": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2853087168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1366.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2853558208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1367.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2853559488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1368.const", - "file_size": 1280 - }, - "onnx::MatMul_18390_11_0_279_12_65_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2853560768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1369.const", - "file_size": 1868800 - }, - "onnx::MatMul_18390_11_0_279_12_65_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2855429568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1370.const", - "file_size": 1868800 - }, - "onnx::MatMul_18391_11_0_280": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2857298368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1371.const", - "file_size": 1884160 - }, - "onnx::MatMul_18392_11_0_281": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2859182528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1372.const", - "file_size": 471040 - }, - "GroupNorm_33_wts_6_0_24": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 2560 - ], - "size_in_bytes": 5120, - "op_tensor_size": 5120, - "offset": 2859653568, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1373.const", - "file_size": 5120 - }, - "NhwcConv_34_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 8622080 - ], - "size_in_bytes": 8622080, - "op_tensor_size": 8622080, - "offset": 2859658688, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1374.const", - "file_size": 8622080 - }, - "GroupNorm_34_wts_6_0_25": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2868280768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1375.const", - "file_size": 2560 - }, - "NhwcConv_35_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 2868283328, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1376.const", - "file_size": 4311040 - }, - "NhwcConv_36_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1003520 - ], - "size_in_bytes": 1003520, - "op_tensor_size": 1003520, - "offset": 2872594368, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1377.const", - "file_size": 1003520 - }, - "GroupNorm_35_wts_6_1_9": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2873597888, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1378.const", - "file_size": 2560 - }, - "onnx::MatMul_18399_11_0_282": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2873600448, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1379.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2874071488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1380.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2874072768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1381.const", - "file_size": 1280 - }, - "Attention_66_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2874074048, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1382.const", - "file_size": 471040 - }, - "Attention_66_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2874545088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1383.const", - "file_size": 471040 - }, - "Attention_66_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2875016128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1384.const", - "file_size": 471040 - }, - "onnx::MatMul_18410_11_0_283": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2875487168, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1385.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2875958208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1386.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2875959488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1387.const", - "file_size": 1280 - }, - "onnx::MatMul_18411": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2875960768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1388.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2876431808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1389.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2876431936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1390.const", - "file_size": 128 - }, - "onnx::MatMul_18421_11_0_284": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2876432064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1391.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2876903104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1392.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2876904384, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1393.const", - "file_size": 1280 - }, - "onnx::MatMul_18422_11_0_285_12_66_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2876905664, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1394.const", - "file_size": 1868800 - }, - "onnx::MatMul_18422_11_0_285_12_66_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2878774464, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1395.const", - "file_size": 1868800 - }, - "onnx::MatMul_18423_11_0_286": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2880643264, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1396.const", - "file_size": 1884160 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2882527424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1397.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2882528704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1398.const", - "file_size": 1280 - }, - "Attention_67_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2882529984, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1399.const", - "file_size": 471040 - }, - "Attention_67_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2883001024, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1400.const", - "file_size": 471040 - }, - "Attention_67_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2883472064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1401.const", - "file_size": 471040 - }, - "onnx::MatMul_18434_11_0_287": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2883943104, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1402.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2884414144, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1403.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2884415424, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1404.const", - "file_size": 1280 - }, - "onnx::MatMul_18435": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2884416704, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1405.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2884887744, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1406.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2884887872, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1407.const", - "file_size": 128 - }, - "onnx::MatMul_18445_11_0_288": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2884888000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1408.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2885359040, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1409.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2885360320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1410.const", - "file_size": 1280 - }, - "onnx::MatMul_18446_11_0_289_12_67_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2885361600, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1411.const", - "file_size": 1868800 - }, - "onnx::MatMul_18446_11_0_289_12_67_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2887230400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1412.const", - "file_size": 1868800 - }, - "onnx::MatMul_18447_11_0_290": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2889099200, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1413.const", - "file_size": 1884160 - }, - "onnx::MatMul_18448_11_0_291": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2890983360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1414.const", - "file_size": 471040 - }, - "GroupNorm_36_wts_6_0_26": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1920 - ], - "size_in_bytes": 3840, - "op_tensor_size": 3840, - "offset": 2891454400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1415.const", - "file_size": 3840 - }, - "NhwcConv_37_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 6425600 - ], - "size_in_bytes": 6425600, - "op_tensor_size": 6425600, - "offset": 2891458240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1416.const", - "file_size": 6425600 - }, - "GroupNorm_37_wts_6_0_27": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2897883840, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1417.const", - "file_size": 2560 - }, - "NhwcConv_38_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4311040 - ], - "size_in_bytes": 4311040, - "op_tensor_size": 4311040, - "offset": 2897886400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1418.const", - "file_size": 4311040 - }, - "NhwcConv_39_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 752640 - ], - "size_in_bytes": 752640, - "op_tensor_size": 752640, - "offset": 2902197440, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1419.const", - "file_size": 752640 - }, - "GroupNorm_38_wts_6_1_10": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2902950080, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1420.const", - "file_size": 2560 - }, - "onnx::MatMul_18455_11_0_292": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2902952640, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1421.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2903423680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1422.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2903424960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1423.const", - "file_size": 1280 - }, - "Attention_68_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2903426240, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1424.const", - "file_size": 471040 - }, - "Attention_68_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2903897280, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1425.const", - "file_size": 471040 - }, - "Attention_68_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2904368320, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1426.const", - "file_size": 471040 - }, - "onnx::MatMul_18466_11_0_293": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2904839360, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1427.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2905310400, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1428.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2905311680, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1429.const", - "file_size": 1280 - }, - "onnx::MatMul_18467": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2905312960, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1430.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2905784000, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1431.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2905784128, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1432.const", - "file_size": 128 - }, - "onnx::MatMul_18477_11_0_294": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2905784256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1433.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2906255296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1434.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2906256576, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1435.const", - "file_size": 1280 - }, - "onnx::MatMul_18478_11_0_295_12_68_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2906257856, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1436.const", - "file_size": 1868800 - }, - "onnx::MatMul_18478_11_0_295_12_68_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2908126656, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1437.const", - "file_size": 1868800 - }, - "onnx::MatMul_18479_11_0_296": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2909995456, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1438.const", - "file_size": 1884160 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2911879616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1439.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2911880896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1440.const", - "file_size": 1280 - }, - "Attention_69_qkv_weight_key": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2911882176, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1441.const", - "file_size": 471040 - }, - "Attention_69_qkv_weight_query": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2912353216, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1442.const", - "file_size": 471040 - }, - "Attention_69_qkv_weight_value": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2912824256, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1443.const", - "file_size": 471040 - }, - "onnx::MatMul_18490_11_0_297": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2913295296, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1444.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2913766336, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1445.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2913767616, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1446.const", - "file_size": 1280 - }, - "onnx::MatMul_18491": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2913768896, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1447.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2914239936, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1448.const", - "file_size": 128 - }, - "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2914240064, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1449.const", - "file_size": 128 - }, - "onnx::MatMul_18501_11_0_298": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2914240192, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1450.const", - "file_size": 471040 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2914711232, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1451.const", - "file_size": 1280 - }, - "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2914712512, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1452.const", - "file_size": 1280 - }, - "onnx::MatMul_18502_11_0_299_12_69_0": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2914713792, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1453.const", - "file_size": 1868800 - }, - "onnx::MatMul_18502_11_0_299_12_69_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1868800 - ], - "size_in_bytes": 1868800, - "op_tensor_size": 1868800, - "offset": 2916582592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1454.const", - "file_size": 1868800 - }, - "onnx::MatMul_18503_11_0_300": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1884160 - ], - "size_in_bytes": 1884160, - "op_tensor_size": 1884160, - "offset": 2918451392, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1455.const", - "file_size": 1884160 - }, - "onnx::MatMul_18504_11_0_301": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 471040 - ], - "size_in_bytes": 471040, - "op_tensor_size": 471040, - "offset": 2920335552, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1456.const", - "file_size": 471040 - }, - "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 128 - ], - "size_in_bytes": 256, - "op_tensor_size": 256, - "offset": 2920806592, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1457.const", - "file_size": 256 - }, - "NhwcConv_40_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 4229120 - ], - "size_in_bytes": 4229120, - "op_tensor_size": 4229120, - "offset": 2920806848, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1458.const", - "file_size": 4229120 - }, - "GroupNorm_39_wts_6_0_28": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1920 - ], - "size_in_bytes": 3840, - "op_tensor_size": 3840, - "offset": 2925035968, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1459.const", - "file_size": 3840 - }, - "NhwcConv_41_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 3171840 - ], - "size_in_bytes": 3171840, - "op_tensor_size": 3171840, - "offset": 2925039808, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1460.const", - "file_size": 3171840 - }, - "GroupNorm_40_wts_6_0_29": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2928211648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1461.const", - "file_size": 1280 - }, - "NhwcConv_42_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 2928212928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1462.const", - "file_size": 1057280 - }, - "NhwcConv_43_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 376320 - ], - "size_in_bytes": 376320, - "op_tensor_size": 376320, - "offset": 2929270208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1463.const", - "file_size": 376320 - }, - "GroupNorm_41_wts_6_0_30": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2929646528, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1464.const", - "file_size": 2560 - }, - "NhwcConv_44_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2114560 - ], - "size_in_bytes": 2114560, - "op_tensor_size": 2114560, - "offset": 2929649088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1465.const", - "file_size": 2114560 - }, - "GroupNorm_42_wts_6_0_31": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2931763648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1466.const", - "file_size": 1280 - }, - "NhwcConv_45_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 2931764928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1467.const", - "file_size": 1057280 - }, - "NhwcConv_46_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 250880 - ], - "size_in_bytes": 250880, - "op_tensor_size": 250880, - "offset": 2932822208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1468.const", - "file_size": 250880 - }, - "GroupNorm_43_wts_6_0_32": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 1280 - ], - "size_in_bytes": 2560, - "op_tensor_size": 2560, - "offset": 2933073088, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1469.const", - "file_size": 2560 - }, - "NhwcConv_47_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 2114560 - ], - "size_in_bytes": 2114560, - "op_tensor_size": 2114560, - "offset": 2933075648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1470.const", - "file_size": 2114560 - }, - "GroupNorm_44_wts_6_0_33": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2935190208, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1471.const", - "file_size": 1280 - }, - "NhwcConv_48_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 1057280 - ], - "size_in_bytes": 1057280, - "op_tensor_size": 1057280, - "offset": 2935191488, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1472.const", - "file_size": 1057280 - }, - "NhwcConv_49_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 250880 - ], - "size_in_bytes": 250880, - "op_tensor_size": 250880, - "offset": 2936248768, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1473.const", - "file_size": 250880 - }, - "GroupNorm_45_wts_6_0_34": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 640 - ], - "size_in_bytes": 1280, - "op_tensor_size": 1280, - "offset": 2936499648, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1474.const", - "file_size": 1280 - }, - "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfloat16", - "shape": [ - 64 - ], - "size_in_bytes": 128, - "op_tensor_size": 128, - "offset": 2936500928, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1475.const", - "file_size": 128 - }, - "NhwcConv_50_weight_NHWC": { - "packed_buffer_label": "const", - "xrt_arg_id": 3, - "dtype": "bfp16ebs8", - "shape": [ - 54400 - ], - "size_in_bytes": 54400, - "op_tensor_size": 54400, - "offset": 2936501056, - "file_name": ".cache/NhwcConv_0-unetconv_inConv_1476.const", - "file_size": 54400 - } - }, - "aux_info": {} +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "NhwcConv_0-/unet/conv_in/Conv", + "type": "SDConv", + "in_args": [ + "Transpose_34_out-sample.out0_1_0" + ], + "const_args": [ + "NhwcConv_0_weight_NHWC" + ], + "out_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "4" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "4" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_SDCastBf2Bfp_NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", + "type": "SDCastBf2Bfp", + "in_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0" + ], + "const_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts" + ], + "out_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_0", + "type": "SDGroupNorm_bfp", + "in_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" + ], + "const_args": [ + "GroupNorm_0_wts_6_0_0" + ], + "out_args": [ + "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/add_embedding/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/Concat_3_output_0.out2_3_1" + ], + "const_args": [ + "unet.add_embedding.linear_1.weight_11_1_0" + ], + "out_args": [ + "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "2816" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2816", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid", + "type": "SDSilu", + "in_args": [ + "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1" + ], + "const_args": [ + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1" + ], + "out_args": [ + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/add_embedding/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1" + ], + "const_args": [ + "unet.add_embedding.linear_2.weight_11_1_1" + ], + "out_args": [ + "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16751" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16752" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16775" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16776" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16807" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16808" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16831" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16832" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16863" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16864" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16887" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16888" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16911" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16912" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16935" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16936" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16959" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16960" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16983" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_16984" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17007" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17008" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17031" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17032" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17055" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17056" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17079" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17080" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17111" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17112" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17135" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17136" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17159" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17160" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17183" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17184" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17207" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17208" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17231" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17232" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17255" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17256" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17279" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17280" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17303" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17304" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17327" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17328" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17359" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17360" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17383" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17384" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17407" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17408" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17431" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17432" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17455" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17456" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17479" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17480" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17503" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17504" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17527" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17528" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17551" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17552" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17575" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17576" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/time_embedding/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/time_proj/Concat_1_output_0.out2_3_0" + ], + "const_args": [ + "unet.time_embedding.linear_1.weight_11_1_2" + ], + "out_args": [ + "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid", + "type": "SDSilu", + "in_args": [ + "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0" + ], + "const_args": [ + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0" + ], + "out_args": [ + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/time_embedding/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0" + ], + "const_args": [ + "unet.time_embedding.linear_2.weight_11_1_3" + ], + "out_args": [ + "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/Add", + "type": "SDAdd", + "in_args": [ + "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2", + "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3" + ], + "const_args": [], + "out_args": [ + "/unet/Add_output_0.out_17_1_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid", + "type": "SDSilu", + "in_args": [ + "/unet/Add_output_0.out_17_1_0" + ], + "const_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2" + ], + "out_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4" + ], + "out_args": [ + "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_1-/unet/down_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0" + ], + "const_args": [ + "NhwcConv_1_weight_NHWC" + ], + "out_args": [ + "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1", + "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_1", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1" + ], + "const_args": [ + "GroupNorm_1_wts_6_0_1" + ], + "out_args": [ + "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_2-/unet/down_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1" + ], + "const_args": [ + "NhwcConv_2_weight_NHWC" + ], + "out_args": [ + "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", + "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_2", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" + ], + "const_args": [ + "GroupNorm_2_wts_6_0_2" + ], + "out_args": [ + "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5" + ], + "out_args": [ + "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_3-/unet/down_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2" + ], + "const_args": [ + "NhwcConv_3_weight_NHWC" + ], + "out_args": [ + "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3", + "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_3", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3" + ], + "const_args": [ + "GroupNorm_3_wts_6_0_3" + ], + "out_args": [ + "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_4-/unet/down_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3" + ], + "const_args": [ + "NhwcConv_4_weight_NHWC" + ], + "out_args": [ + "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.0/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2", + "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_5-/unet/down_blocks.0/downsamplers.0/conv/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" + ], + "const_args": [ + "NhwcConv_5_weight_NHWC" + ], + "out_args": [ + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_4", + "type": "SDGroupNorm_bfp", + "in_args": [ + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" + ], + "const_args": [ + "GroupNorm_4_wts_6_0_4" + ], + "out_args": [ + "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6" + ], + "out_args": [ + "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_6-/unet/down_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4" + ], + "const_args": [ + "NhwcConv_6_weight_NHWC" + ], + "out_args": [ + "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7", + "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_5", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5" + ], + "const_args": [ + "GroupNorm_5_wts_6_0_5" + ], + "out_args": [ + "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7" + ], + "out_args": [ + "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8" + ], + "out_args": [ + "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9" + ], + "out_args": [ + "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10" + ], + "out_args": [ + "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11" + ], + "out_args": [ + "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17611" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17612" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17635" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17636" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17659" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17660" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17683" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17684" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17707" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17708" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17731" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17732" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17755" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17756" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17779" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17780" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17803" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17804" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17827" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17828" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17859" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17860" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17883" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17884" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17907" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17908" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17931" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17932" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17955" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17956" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17979" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_17980" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18003" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18004" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18027" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18028" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18051" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18052" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18075" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18076" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18107" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18108" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18131" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18132" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18155" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18156" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18179" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18180" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18203" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18204" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18227" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18228" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18251" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18252" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18275" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18276" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18299" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18300" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18323" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18324" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12" + ], + "out_args": [ + "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13" + ], + "out_args": [ + "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14" + ], + "out_args": [ + "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18356" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18357" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18380" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18381" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18412" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18413" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18436" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18437" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18468" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18469" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18492" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_18493" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "2048" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15" + ], + "out_args": [ + "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16" + ], + "out_args": [ + "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17" + ], + "out_args": [ + "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18" + ], + "out_args": [ + "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19" + ], + "out_args": [ + "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2" + ], + "const_args": [ + "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20" + ], + "out_args": [ + "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_7-/unet/down_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5" + ], + "const_args": [ + "NhwcConv_7_weight_NHWC" + ], + "out_args": [ + "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_8-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" + ], + "const_args": [ + "NhwcConv_8_weight_NHWC" + ], + "out_args": [ + "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6", + "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_6", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" + ], + "const_args": [ + "GroupNorm_6_wts_6_1_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6" + ], + "const_args": [ + "onnx::MatMul_16738_11_0_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" + ], + "const_args": [ + "Attention_0_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" + ], + "const_args": [ + "Attention_0_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0" + ], + "const_args": [ + "Attention_0_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_0", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0" + ], + "const_args": [ + "onnx::MatMul_16749_11_0_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235", + "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1" + ], + "const_args": [ + "onnx::MatMul_16750" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_SDCastBfp2Bf_/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_0", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_SDCastBf2Bfp_/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236" + ], + "const_args": [ + "onnx::MatMul_16760_11_0_2" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_0_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" + ], + "const_args": [ + "onnx::MatMul_16761_11_0_3_12_0_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_0_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2" + ], + "const_args": [ + "onnx::MatMul_16761_11_0_3_12_0_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0" + ], + "const_args": [ + "onnx::MatMul_16762_11_0_4" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" + ], + "const_args": [ + "Attention_1_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" + ], + "const_args": [ + "Attention_1_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3" + ], + "const_args": [ + "Attention_1_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_1", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1" + ], + "const_args": [ + "onnx::MatMul_16773_11_0_5" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4" + ], + "const_args": [ + "onnx::MatMul_16774" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_SDCastBfp2Bf_/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_1", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_SDCastBf2Bfp_/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244" + ], + "const_args": [ + "onnx::MatMul_16784_11_0_6" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" + ], + "const_args": [ + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_1_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" + ], + "const_args": [ + "onnx::MatMul_16785_11_0_7_12_1_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_1_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5" + ], + "const_args": [ + "onnx::MatMul_16785_11_0_7_12_1_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1" + ], + "const_args": [ + "onnx::MatMul_16786_11_0_8" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12" + ], + "const_args": [ + "onnx::MatMul_16787_11_0_9" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248", + "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_7", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" + ], + "const_args": [ + "GroupNorm_7_wts_6_0_6" + ], + "out_args": [ + "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_9-/unet/down_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7" + ], + "const_args": [ + "NhwcConv_9_weight_NHWC" + ], + "out_args": [ + "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9", + "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_8", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14" + ], + "const_args": [ + "GroupNorm_8_wts_6_0_7" + ], + "out_args": [ + "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_10-/unet/down_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8" + ], + "const_args": [ + "NhwcConv_10_weight_NHWC" + ], + "out_args": [ + "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13", + "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_9", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" + ], + "const_args": [ + "GroupNorm_9_wts_6_1_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9" + ], + "const_args": [ + "onnx::MatMul_16794_11_0_10" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" + ], + "const_args": [ + "Attention_2_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" + ], + "const_args": [ + "Attention_2_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6" + ], + "const_args": [ + "Attention_2_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_2", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2" + ], + "const_args": [ + "onnx::MatMul_16805_11_0_11" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253", + "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7" + ], + "const_args": [ + "onnx::MatMul_16806" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_SDCastBfp2Bf_/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_2", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_SDCastBf2Bfp_/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254" + ], + "const_args": [ + "onnx::MatMul_16816_11_0_12" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_2_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" + ], + "const_args": [ + "onnx::MatMul_16817_11_0_13_12_2_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_2_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8" + ], + "const_args": [ + "onnx::MatMul_16817_11_0_13_12_2_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2" + ], + "const_args": [ + "onnx::MatMul_16818_11_0_14" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" + ], + "const_args": [ + "Attention_3_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" + ], + "const_args": [ + "Attention_3_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9" + ], + "const_args": [ + "Attention_3_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_3", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3" + ], + "const_args": [ + "onnx::MatMul_16829_11_0_15" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10" + ], + "const_args": [ + "onnx::MatMul_16830" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_SDCastBfp2Bf_/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_3", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_SDCastBf2Bfp_/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7" + ], + "const_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262" + ], + "const_args": [ + "onnx::MatMul_16840_11_0_16" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" + ], + "const_args": [ + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_3_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" + ], + "const_args": [ + "onnx::MatMul_16841_11_0_17_12_3_0" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_3_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11" + ], + "const_args": [ + "onnx::MatMul_16841_11_0_17_12_3_1" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3" + ], + "const_args": [ + "onnx::MatMul_16842_11_0_18" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21" + ], + "const_args": [ + "onnx::MatMul_16843_11_0_19" + ], + "out_args": [ + "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.1/attentions.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266", + "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_11-/unet/down_blocks.1/downsamplers.0/conv/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" + ], + "const_args": [ + "NhwcConv_11_weight_NHWC" + ], + "out_args": [ + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_10", + "type": "SDGroupNorm_bfp", + "in_args": [ + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" + ], + "const_args": [ + "GroupNorm_10_wts_6_0_8" + ], + "out_args": [ + "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_12-/unet/down_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10" + ], + "const_args": [ + "NhwcConv_12_weight_NHWC" + ], + "out_args": [ + "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13", + "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_11", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23" + ], + "const_args": [ + "GroupNorm_11_wts_6_0_9" + ], + "out_args": [ + "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_13-/unet/down_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11" + ], + "const_args": [ + "NhwcConv_13_weight_NHWC" + ], + "out_args": [ + "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_14-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" + ], + "const_args": [ + "NhwcConv_14_weight_NHWC" + ], + "out_args": [ + "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12", + "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_12", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" + ], + "const_args": [ + "GroupNorm_12_wts_6_1_2" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12" + ], + "const_args": [ + "onnx::MatMul_16850_11_0_20" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" + ], + "const_args": [ + "Attention_4_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" + ], + "const_args": [ + "Attention_4_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12" + ], + "const_args": [ + "Attention_4_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_4", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4" + ], + "const_args": [ + "onnx::MatMul_16861_11_0_21" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271", + "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13" + ], + "const_args": [ + "onnx::MatMul_16862" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_4", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272" + ], + "const_args": [ + "onnx::MatMul_16872_11_0_22" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_4_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" + ], + "const_args": [ + "onnx::MatMul_16873_11_0_23_12_4_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_4_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14" + ], + "const_args": [ + "onnx::MatMul_16873_11_0_23_12_4_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4" + ], + "const_args": [ + "onnx::MatMul_16874_11_0_24" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" + ], + "const_args": [ + "Attention_5_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" + ], + "const_args": [ + "Attention_5_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15" + ], + "const_args": [ + "Attention_5_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_5", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5" + ], + "const_args": [ + "onnx::MatMul_16885_11_0_25" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16" + ], + "const_args": [ + "onnx::MatMul_16886" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_5", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280" + ], + "const_args": [ + "onnx::MatMul_16896_11_0_26" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_5_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" + ], + "const_args": [ + "onnx::MatMul_16897_11_0_27_12_5_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_5_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17" + ], + "const_args": [ + "onnx::MatMul_16897_11_0_27_12_5_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5" + ], + "const_args": [ + "onnx::MatMul_16898_11_0_28" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" + ], + "const_args": [ + "Attention_6_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" + ], + "const_args": [ + "Attention_6_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18" + ], + "const_args": [ + "Attention_6_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_6", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6" + ], + "const_args": [ + "onnx::MatMul_16909_11_0_29" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19" + ], + "const_args": [ + "onnx::MatMul_16910" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_6", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288" + ], + "const_args": [ + "onnx::MatMul_16920_11_0_30" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_6_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" + ], + "const_args": [ + "onnx::MatMul_16921_11_0_31_12_6_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_6_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20" + ], + "const_args": [ + "onnx::MatMul_16921_11_0_31_12_6_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6" + ], + "const_args": [ + "onnx::MatMul_16922_11_0_32" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" + ], + "const_args": [ + "Attention_7_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" + ], + "const_args": [ + "Attention_7_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21" + ], + "const_args": [ + "Attention_7_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_7", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7" + ], + "const_args": [ + "onnx::MatMul_16933_11_0_33" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22" + ], + "const_args": [ + "onnx::MatMul_16934" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_7", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296" + ], + "const_args": [ + "onnx::MatMul_16944_11_0_34" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_7_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" + ], + "const_args": [ + "onnx::MatMul_16945_11_0_35_12_7_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_7_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23" + ], + "const_args": [ + "onnx::MatMul_16945_11_0_35_12_7_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7" + ], + "const_args": [ + "onnx::MatMul_16946_11_0_36" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" + ], + "const_args": [ + "Attention_8_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" + ], + "const_args": [ + "Attention_8_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24" + ], + "const_args": [ + "Attention_8_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_8", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8" + ], + "const_args": [ + "onnx::MatMul_16957_11_0_37" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25" + ], + "const_args": [ + "onnx::MatMul_16958" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_8", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304" + ], + "const_args": [ + "onnx::MatMul_16968_11_0_38" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_8_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" + ], + "const_args": [ + "onnx::MatMul_16969_11_0_39_12_8_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_8_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26" + ], + "const_args": [ + "onnx::MatMul_16969_11_0_39_12_8_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8" + ], + "const_args": [ + "onnx::MatMul_16970_11_0_40" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" + ], + "const_args": [ + "Attention_9_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" + ], + "const_args": [ + "Attention_9_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27" + ], + "const_args": [ + "Attention_9_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_9", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9" + ], + "const_args": [ + "onnx::MatMul_16981_11_0_41" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28" + ], + "const_args": [ + "onnx::MatMul_16982" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_9", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312" + ], + "const_args": [ + "onnx::MatMul_16992_11_0_42" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_9_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" + ], + "const_args": [ + "onnx::MatMul_16993_11_0_43_12_9_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_9_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29" + ], + "const_args": [ + "onnx::MatMul_16993_11_0_43_12_9_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9" + ], + "const_args": [ + "onnx::MatMul_16994_11_0_44" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" + ], + "const_args": [ + "Attention_10_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" + ], + "const_args": [ + "Attention_10_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30" + ], + "const_args": [ + "Attention_10_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_10", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10" + ], + "const_args": [ + "onnx::MatMul_17005_11_0_45" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31" + ], + "const_args": [ + "onnx::MatMul_17006" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_10", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320" + ], + "const_args": [ + "onnx::MatMul_17016_11_0_46" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_10_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" + ], + "const_args": [ + "onnx::MatMul_17017_11_0_47_12_10_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_10_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32" + ], + "const_args": [ + "onnx::MatMul_17017_11_0_47_12_10_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10" + ], + "const_args": [ + "onnx::MatMul_17018_11_0_48" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" + ], + "const_args": [ + "Attention_11_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" + ], + "const_args": [ + "Attention_11_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33" + ], + "const_args": [ + "Attention_11_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_11", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11" + ], + "const_args": [ + "onnx::MatMul_17029_11_0_49" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34" + ], + "const_args": [ + "onnx::MatMul_17030" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_11", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328" + ], + "const_args": [ + "onnx::MatMul_17040_11_0_50" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_11_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" + ], + "const_args": [ + "onnx::MatMul_17041_11_0_51_12_11_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_11_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35" + ], + "const_args": [ + "onnx::MatMul_17041_11_0_51_12_11_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11" + ], + "const_args": [ + "onnx::MatMul_17042_11_0_52" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" + ], + "const_args": [ + "Attention_12_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" + ], + "const_args": [ + "Attention_12_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36" + ], + "const_args": [ + "Attention_12_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_12", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12" + ], + "const_args": [ + "onnx::MatMul_17053_11_0_53" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37" + ], + "const_args": [ + "onnx::MatMul_17054" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_12", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336" + ], + "const_args": [ + "onnx::MatMul_17064_11_0_54" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_12_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" + ], + "const_args": [ + "onnx::MatMul_17065_11_0_55_12_12_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_12_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38" + ], + "const_args": [ + "onnx::MatMul_17065_11_0_55_12_12_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12" + ], + "const_args": [ + "onnx::MatMul_17066_11_0_56" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" + ], + "const_args": [ + "Attention_13_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" + ], + "const_args": [ + "Attention_13_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39" + ], + "const_args": [ + "Attention_13_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_13", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13" + ], + "const_args": [ + "onnx::MatMul_17077_11_0_57" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40" + ], + "const_args": [ + "onnx::MatMul_17078" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_SDCastBfp2Bf_/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_13", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_SDCastBf2Bfp_/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344" + ], + "const_args": [ + "onnx::MatMul_17088_11_0_58" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" + ], + "const_args": [ + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_13_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" + ], + "const_args": [ + "onnx::MatMul_17089_11_0_59_12_13_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_13_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41" + ], + "const_args": [ + "onnx::MatMul_17089_11_0_59_12_13_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13" + ], + "const_args": [ + "onnx::MatMul_17090_11_0_60" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54" + ], + "const_args": [ + "onnx::MatMul_17091_11_0_61" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348", + "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_13", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" + ], + "const_args": [ + "GroupNorm_13_wts_6_0_10" + ], + "out_args": [ + "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_15-/unet/down_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13" + ], + "const_args": [ + "NhwcConv_15_weight_NHWC" + ], + "out_args": [ + "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15", + "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_14", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56" + ], + "const_args": [ + "GroupNorm_14_wts_6_0_11" + ], + "out_args": [ + "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_16-/unet/down_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14" + ], + "const_args": [ + "NhwcConv_16_weight_NHWC" + ], + "out_args": [ + "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55", + "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_15", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" + ], + "const_args": [ + "GroupNorm_15_wts_6_1_3" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15" + ], + "const_args": [ + "onnx::MatMul_17098_11_0_62" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" + ], + "const_args": [ + "Attention_14_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" + ], + "const_args": [ + "Attention_14_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42" + ], + "const_args": [ + "Attention_14_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_14", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14" + ], + "const_args": [ + "onnx::MatMul_17109_11_0_63" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353", + "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43" + ], + "const_args": [ + "onnx::MatMul_17110" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_14", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354" + ], + "const_args": [ + "onnx::MatMul_17120_11_0_64" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_14_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" + ], + "const_args": [ + "onnx::MatMul_17121_11_0_65_12_14_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_14_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44" + ], + "const_args": [ + "onnx::MatMul_17121_11_0_65_12_14_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14" + ], + "const_args": [ + "onnx::MatMul_17122_11_0_66" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" + ], + "const_args": [ + "Attention_15_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" + ], + "const_args": [ + "Attention_15_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45" + ], + "const_args": [ + "Attention_15_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_15", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15" + ], + "const_args": [ + "onnx::MatMul_17133_11_0_67" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46" + ], + "const_args": [ + "onnx::MatMul_17134" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_15", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362" + ], + "const_args": [ + "onnx::MatMul_17144_11_0_68" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_15_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" + ], + "const_args": [ + "onnx::MatMul_17145_11_0_69_12_15_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_15_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47" + ], + "const_args": [ + "onnx::MatMul_17145_11_0_69_12_15_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15" + ], + "const_args": [ + "onnx::MatMul_17146_11_0_70" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" + ], + "const_args": [ + "Attention_16_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" + ], + "const_args": [ + "Attention_16_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48" + ], + "const_args": [ + "Attention_16_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_16", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16" + ], + "const_args": [ + "onnx::MatMul_17157_11_0_71" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49" + ], + "const_args": [ + "onnx::MatMul_17158" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_16", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370" + ], + "const_args": [ + "onnx::MatMul_17168_11_0_72" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_16_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" + ], + "const_args": [ + "onnx::MatMul_17169_11_0_73_12_16_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_16_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50" + ], + "const_args": [ + "onnx::MatMul_17169_11_0_73_12_16_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16" + ], + "const_args": [ + "onnx::MatMul_17170_11_0_74" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" + ], + "const_args": [ + "Attention_17_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" + ], + "const_args": [ + "Attention_17_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51" + ], + "const_args": [ + "Attention_17_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_17", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17" + ], + "const_args": [ + "onnx::MatMul_17181_11_0_75" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52" + ], + "const_args": [ + "onnx::MatMul_17182" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_17", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378" + ], + "const_args": [ + "onnx::MatMul_17192_11_0_76" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_17_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" + ], + "const_args": [ + "onnx::MatMul_17193_11_0_77_12_17_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_17_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53" + ], + "const_args": [ + "onnx::MatMul_17193_11_0_77_12_17_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17" + ], + "const_args": [ + "onnx::MatMul_17194_11_0_78" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" + ], + "const_args": [ + "Attention_18_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" + ], + "const_args": [ + "Attention_18_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54" + ], + "const_args": [ + "Attention_18_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_18", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18" + ], + "const_args": [ + "onnx::MatMul_17205_11_0_79" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55" + ], + "const_args": [ + "onnx::MatMul_17206" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_18", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386" + ], + "const_args": [ + "onnx::MatMul_17216_11_0_80" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_18_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" + ], + "const_args": [ + "onnx::MatMul_17217_11_0_81_12_18_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_18_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56" + ], + "const_args": [ + "onnx::MatMul_17217_11_0_81_12_18_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18" + ], + "const_args": [ + "onnx::MatMul_17218_11_0_82" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" + ], + "const_args": [ + "Attention_19_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" + ], + "const_args": [ + "Attention_19_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57" + ], + "const_args": [ + "Attention_19_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_19", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19" + ], + "const_args": [ + "onnx::MatMul_17229_11_0_83" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58" + ], + "const_args": [ + "onnx::MatMul_17230" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_19", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394" + ], + "const_args": [ + "onnx::MatMul_17240_11_0_84" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_19_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" + ], + "const_args": [ + "onnx::MatMul_17241_11_0_85_12_19_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_19_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59" + ], + "const_args": [ + "onnx::MatMul_17241_11_0_85_12_19_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19" + ], + "const_args": [ + "onnx::MatMul_17242_11_0_86" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" + ], + "const_args": [ + "Attention_20_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" + ], + "const_args": [ + "Attention_20_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60" + ], + "const_args": [ + "Attention_20_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_20", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20" + ], + "const_args": [ + "onnx::MatMul_17253_11_0_87" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61" + ], + "const_args": [ + "onnx::MatMul_17254" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_20", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402" + ], + "const_args": [ + "onnx::MatMul_17264_11_0_88" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_20_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" + ], + "const_args": [ + "onnx::MatMul_17265_11_0_89_12_20_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_20_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62" + ], + "const_args": [ + "onnx::MatMul_17265_11_0_89_12_20_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20" + ], + "const_args": [ + "onnx::MatMul_17266_11_0_90" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" + ], + "const_args": [ + "Attention_21_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" + ], + "const_args": [ + "Attention_21_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63" + ], + "const_args": [ + "Attention_21_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_21", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21" + ], + "const_args": [ + "onnx::MatMul_17277_11_0_91" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64" + ], + "const_args": [ + "onnx::MatMul_17278" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_21", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410" + ], + "const_args": [ + "onnx::MatMul_17288_11_0_92" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_21_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" + ], + "const_args": [ + "onnx::MatMul_17289_11_0_93_12_21_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_21_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65" + ], + "const_args": [ + "onnx::MatMul_17289_11_0_93_12_21_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21" + ], + "const_args": [ + "onnx::MatMul_17290_11_0_94" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" + ], + "const_args": [ + "Attention_22_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" + ], + "const_args": [ + "Attention_22_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66" + ], + "const_args": [ + "Attention_22_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_22", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22" + ], + "const_args": [ + "onnx::MatMul_17301_11_0_95" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67" + ], + "const_args": [ + "onnx::MatMul_17302" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_22", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418" + ], + "const_args": [ + "onnx::MatMul_17312_11_0_96" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_22_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" + ], + "const_args": [ + "onnx::MatMul_17313_11_0_97_12_22_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_22_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68" + ], + "const_args": [ + "onnx::MatMul_17313_11_0_97_12_22_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22" + ], + "const_args": [ + "onnx::MatMul_17314_11_0_98" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" + ], + "const_args": [ + "Attention_23_qkv_weight_key" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" + ], + "const_args": [ + "Attention_23_qkv_weight_query" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69" + ], + "const_args": [ + "Attention_23_qkv_weight_value" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_23", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23" + ], + "const_args": [ + "onnx::MatMul_17325_11_0_99" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70" + ], + "const_args": [ + "onnx::MatMul_17326" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_SDCastBfp2Bf_/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_23", + "type": "SDFlatMHA", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_SDCastBf2Bfp_/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47" + ], + "const_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426" + ], + "const_args": [ + "onnx::MatMul_17336_11_0_100" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" + ], + "const_args": [ + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_23_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" + ], + "const_args": [ + "onnx::MatMul_17337_11_0_101_12_23_0" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_23_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71" + ], + "const_args": [ + "onnx::MatMul_17337_11_0_101_12_23_1" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23" + ], + "const_args": [ + "onnx::MatMul_17338_11_0_102" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87" + ], + "const_args": [ + "onnx::MatMul_17339_11_0_103" + ], + "out_args": [ + "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/down_blocks.2/attentions.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430", + "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57" + ], + "const_args": [], + "out_args": [ + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_16", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" + ], + "const_args": [ + "GroupNorm_16_wts_6_0_12" + ], + "out_args": [ + "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_17-/unet/mid_block/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16" + ], + "const_args": [ + "NhwcConv_17_weight_NHWC" + ], + "out_args": [ + "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17", + "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_17", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89" + ], + "const_args": [ + "GroupNorm_17_wts_6_0_13" + ], + "out_args": [ + "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_18-/unet/mid_block/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17" + ], + "const_args": [ + "NhwcConv_18_weight_NHWC" + ], + "out_args": [ + "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88", + "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_18", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" + ], + "const_args": [ + "GroupNorm_18_wts_6_1_4" + ], + "out_args": [ + "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18" + ], + "const_args": [ + "onnx::MatMul_17346_11_0_104" + ], + "out_args": [ + "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" + ], + "const_args": [ + "Attention_24_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" + ], + "const_args": [ + "Attention_24_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72" + ], + "const_args": [ + "Attention_24_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_24", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24" + ], + "const_args": [ + "onnx::MatMul_17357_11_0_105" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435", + "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73" + ], + "const_args": [ + "onnx::MatMul_17358" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_24", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436" + ], + "const_args": [ + "onnx::MatMul_17368_11_0_106" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_24_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" + ], + "const_args": [ + "onnx::MatMul_17369_11_0_107_12_24_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_24_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74" + ], + "const_args": [ + "onnx::MatMul_17369_11_0_107_12_24_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438", + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24" + ], + "const_args": [ + "onnx::MatMul_17370_11_0_108" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" + ], + "const_args": [ + "Attention_25_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" + ], + "const_args": [ + "Attention_25_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75" + ], + "const_args": [ + "Attention_25_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_25", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25" + ], + "const_args": [ + "onnx::MatMul_17381_11_0_109" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76" + ], + "const_args": [ + "onnx::MatMul_17382" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_25", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444" + ], + "const_args": [ + "onnx::MatMul_17392_11_0_110" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_25_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" + ], + "const_args": [ + "onnx::MatMul_17393_11_0_111_12_25_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_25_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77" + ], + "const_args": [ + "onnx::MatMul_17393_11_0_111_12_25_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446", + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25" + ], + "const_args": [ + "onnx::MatMul_17394_11_0_112" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" + ], + "const_args": [ + "Attention_26_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" + ], + "const_args": [ + "Attention_26_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78" + ], + "const_args": [ + "Attention_26_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_26", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26" + ], + "const_args": [ + "onnx::MatMul_17405_11_0_113" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79" + ], + "const_args": [ + "onnx::MatMul_17406" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_26", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452" + ], + "const_args": [ + "onnx::MatMul_17416_11_0_114" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_26_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" + ], + "const_args": [ + "onnx::MatMul_17417_11_0_115_12_26_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_26_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80" + ], + "const_args": [ + "onnx::MatMul_17417_11_0_115_12_26_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454", + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26" + ], + "const_args": [ + "onnx::MatMul_17418_11_0_116" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" + ], + "const_args": [ + "Attention_27_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" + ], + "const_args": [ + "Attention_27_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81" + ], + "const_args": [ + "Attention_27_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_27", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27" + ], + "const_args": [ + "onnx::MatMul_17429_11_0_117" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82" + ], + "const_args": [ + "onnx::MatMul_17430" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_27", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460" + ], + "const_args": [ + "onnx::MatMul_17440_11_0_118" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_27_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" + ], + "const_args": [ + "onnx::MatMul_17441_11_0_119_12_27_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_27_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83" + ], + "const_args": [ + "onnx::MatMul_17441_11_0_119_12_27_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462", + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27" + ], + "const_args": [ + "onnx::MatMul_17442_11_0_120" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" + ], + "const_args": [ + "Attention_28_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" + ], + "const_args": [ + "Attention_28_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84" + ], + "const_args": [ + "Attention_28_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_28", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28" + ], + "const_args": [ + "onnx::MatMul_17453_11_0_121" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85" + ], + "const_args": [ + "onnx::MatMul_17454" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_28", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468" + ], + "const_args": [ + "onnx::MatMul_17464_11_0_122" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_28_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" + ], + "const_args": [ + "onnx::MatMul_17465_11_0_123_12_28_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_28_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86" + ], + "const_args": [ + "onnx::MatMul_17465_11_0_123_12_28_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470", + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28" + ], + "const_args": [ + "onnx::MatMul_17466_11_0_124" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" + ], + "const_args": [ + "Attention_29_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" + ], + "const_args": [ + "Attention_29_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87" + ], + "const_args": [ + "Attention_29_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_29", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29" + ], + "const_args": [ + "onnx::MatMul_17477_11_0_125" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88" + ], + "const_args": [ + "onnx::MatMul_17478" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_29", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476" + ], + "const_args": [ + "onnx::MatMul_17488_11_0_126" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_29_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" + ], + "const_args": [ + "onnx::MatMul_17489_11_0_127_12_29_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_29_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89" + ], + "const_args": [ + "onnx::MatMul_17489_11_0_127_12_29_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478", + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29" + ], + "const_args": [ + "onnx::MatMul_17490_11_0_128" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" + ], + "const_args": [ + "Attention_30_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" + ], + "const_args": [ + "Attention_30_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90" + ], + "const_args": [ + "Attention_30_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_30", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30" + ], + "const_args": [ + "onnx::MatMul_17501_11_0_129" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91" + ], + "const_args": [ + "onnx::MatMul_17502" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_30", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484" + ], + "const_args": [ + "onnx::MatMul_17512_11_0_130" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_30_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" + ], + "const_args": [ + "onnx::MatMul_17513_11_0_131_12_30_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_30_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92" + ], + "const_args": [ + "onnx::MatMul_17513_11_0_131_12_30_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486", + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30" + ], + "const_args": [ + "onnx::MatMul_17514_11_0_132" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" + ], + "const_args": [ + "Attention_31_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" + ], + "const_args": [ + "Attention_31_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93" + ], + "const_args": [ + "Attention_31_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_31", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31" + ], + "const_args": [ + "onnx::MatMul_17525_11_0_133" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94" + ], + "const_args": [ + "onnx::MatMul_17526" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_31", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492" + ], + "const_args": [ + "onnx::MatMul_17536_11_0_134" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_31_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" + ], + "const_args": [ + "onnx::MatMul_17537_11_0_135_12_31_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_31_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95" + ], + "const_args": [ + "onnx::MatMul_17537_11_0_135_12_31_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494", + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31" + ], + "const_args": [ + "onnx::MatMul_17538_11_0_136" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" + ], + "const_args": [ + "Attention_32_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" + ], + "const_args": [ + "Attention_32_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96" + ], + "const_args": [ + "Attention_32_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_32", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32" + ], + "const_args": [ + "onnx::MatMul_17549_11_0_137" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97" + ], + "const_args": [ + "onnx::MatMul_17550" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_32", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500" + ], + "const_args": [ + "onnx::MatMul_17560_11_0_138" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_32_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" + ], + "const_args": [ + "onnx::MatMul_17561_11_0_139_12_32_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_32_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98" + ], + "const_args": [ + "onnx::MatMul_17561_11_0_139_12_32_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502", + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32" + ], + "const_args": [ + "onnx::MatMul_17562_11_0_140" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" + ], + "const_args": [ + "Attention_33_qkv_weight_key" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" + ], + "const_args": [ + "Attention_33_qkv_weight_query" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99" + ], + "const_args": [ + "Attention_33_qkv_weight_value" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_33", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33" + ], + "const_args": [ + "onnx::MatMul_17573_11_0_141" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100" + ], + "const_args": [ + "onnx::MatMul_17574" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_SDCastBfp2Bf_/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_33", + "type": "SDFlatMHA", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_SDCastBf2Bfp_/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67" + ], + "const_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508" + ], + "const_args": [ + "onnx::MatMul_17584_11_0_142" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508", + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" + ], + "const_args": [ + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_33_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" + ], + "const_args": [ + "onnx::MatMul_17585_11_0_143_12_33_0" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_33_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101" + ], + "const_args": [ + "onnx::MatMul_17585_11_0_143_12_33_1" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510", + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33" + ], + "const_args": [ + "onnx::MatMul_17586_11_0_144" + ], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511", + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120" + ], + "const_args": [ + "onnx::MatMul_17587_11_0_145" + ], + "out_args": [ + "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/attentions.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512", + "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_19", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121" + ], + "const_args": [ + "GroupNorm_19_wts_6_0_14" + ], + "out_args": [ + "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_19-/unet/mid_block/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19" + ], + "const_args": [ + "NhwcConv_19_weight_NHWC" + ], + "out_args": [ + "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19", + "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_20", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122" + ], + "const_args": [ + "GroupNorm_20_wts_6_0_15" + ], + "out_args": [ + "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_20-/unet/mid_block/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20" + ], + "const_args": [ + "NhwcConv_20_weight_NHWC" + ], + "out_args": [ + "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/mid_block/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121", + "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20" + ], + "const_args": [], + "out_args": [ + "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/Concat", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123", + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_21", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" + ], + "const_args": [ + "GroupNorm_21_wts_6_0_16" + ], + "out_args": [ + "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_21-/unet/up_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21" + ], + "const_args": [ + "NhwcConv_21_weight_NHWC" + ], + "out_args": [ + "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22", + "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_22", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124" + ], + "const_args": [ + "GroupNorm_22_wts_6_0_17" + ], + "out_args": [ + "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_22-/unet/up_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22" + ], + "const_args": [ + "NhwcConv_22_weight_NHWC" + ], + "out_args": [ + "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_23-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0" + ], + "const_args": [ + "NhwcConv_23_weight_NHWC" + ], + "out_args": [ + "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21", + "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_23", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" + ], + "const_args": [ + "GroupNorm_23_wts_6_1_5" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23" + ], + "const_args": [ + "onnx::MatMul_17598_11_0_146" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" + ], + "const_args": [ + "Attention_34_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" + ], + "const_args": [ + "Attention_34_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102" + ], + "const_args": [ + "Attention_34_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_34", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34" + ], + "const_args": [ + "onnx::MatMul_17609_11_0_147" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517", + "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103" + ], + "const_args": [ + "onnx::MatMul_17610" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_34", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518" + ], + "const_args": [ + "onnx::MatMul_17620_11_0_148" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_34_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" + ], + "const_args": [ + "onnx::MatMul_17621_11_0_149_12_34_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_34_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104" + ], + "const_args": [ + "onnx::MatMul_17621_11_0_149_12_34_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34" + ], + "const_args": [ + "onnx::MatMul_17622_11_0_150" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" + ], + "const_args": [ + "Attention_35_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" + ], + "const_args": [ + "Attention_35_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105" + ], + "const_args": [ + "Attention_35_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_35", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35" + ], + "const_args": [ + "onnx::MatMul_17633_11_0_151" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106" + ], + "const_args": [ + "onnx::MatMul_17634" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_35", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526" + ], + "const_args": [ + "onnx::MatMul_17644_11_0_152" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_35_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" + ], + "const_args": [ + "onnx::MatMul_17645_11_0_153_12_35_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_35_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107" + ], + "const_args": [ + "onnx::MatMul_17645_11_0_153_12_35_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35" + ], + "const_args": [ + "onnx::MatMul_17646_11_0_154" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" + ], + "const_args": [ + "Attention_36_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" + ], + "const_args": [ + "Attention_36_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108" + ], + "const_args": [ + "Attention_36_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_36", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36" + ], + "const_args": [ + "onnx::MatMul_17657_11_0_155" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109" + ], + "const_args": [ + "onnx::MatMul_17658" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_36", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534" + ], + "const_args": [ + "onnx::MatMul_17668_11_0_156" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_36_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" + ], + "const_args": [ + "onnx::MatMul_17669_11_0_157_12_36_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_36_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110" + ], + "const_args": [ + "onnx::MatMul_17669_11_0_157_12_36_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36" + ], + "const_args": [ + "onnx::MatMul_17670_11_0_158" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" + ], + "const_args": [ + "Attention_37_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" + ], + "const_args": [ + "Attention_37_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111" + ], + "const_args": [ + "Attention_37_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_37", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37" + ], + "const_args": [ + "onnx::MatMul_17681_11_0_159" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112" + ], + "const_args": [ + "onnx::MatMul_17682" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_37", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542" + ], + "const_args": [ + "onnx::MatMul_17692_11_0_160" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_37_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" + ], + "const_args": [ + "onnx::MatMul_17693_11_0_161_12_37_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_37_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113" + ], + "const_args": [ + "onnx::MatMul_17693_11_0_161_12_37_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37" + ], + "const_args": [ + "onnx::MatMul_17694_11_0_162" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" + ], + "const_args": [ + "Attention_38_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" + ], + "const_args": [ + "Attention_38_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114" + ], + "const_args": [ + "Attention_38_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_38", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38" + ], + "const_args": [ + "onnx::MatMul_17705_11_0_163" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115" + ], + "const_args": [ + "onnx::MatMul_17706" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_38", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550" + ], + "const_args": [ + "onnx::MatMul_17716_11_0_164" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_38_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" + ], + "const_args": [ + "onnx::MatMul_17717_11_0_165_12_38_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_38_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116" + ], + "const_args": [ + "onnx::MatMul_17717_11_0_165_12_38_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38" + ], + "const_args": [ + "onnx::MatMul_17718_11_0_166" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" + ], + "const_args": [ + "Attention_39_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" + ], + "const_args": [ + "Attention_39_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117" + ], + "const_args": [ + "Attention_39_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_39", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39" + ], + "const_args": [ + "onnx::MatMul_17729_11_0_167" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118" + ], + "const_args": [ + "onnx::MatMul_17730" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_39", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558" + ], + "const_args": [ + "onnx::MatMul_17740_11_0_168" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_39_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" + ], + "const_args": [ + "onnx::MatMul_17741_11_0_169_12_39_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_39_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119" + ], + "const_args": [ + "onnx::MatMul_17741_11_0_169_12_39_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39" + ], + "const_args": [ + "onnx::MatMul_17742_11_0_170" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" + ], + "const_args": [ + "Attention_40_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" + ], + "const_args": [ + "Attention_40_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120" + ], + "const_args": [ + "Attention_40_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_40", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40" + ], + "const_args": [ + "onnx::MatMul_17753_11_0_171" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121" + ], + "const_args": [ + "onnx::MatMul_17754" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_40", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566" + ], + "const_args": [ + "onnx::MatMul_17764_11_0_172" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_40_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" + ], + "const_args": [ + "onnx::MatMul_17765_11_0_173_12_40_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_40_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122" + ], + "const_args": [ + "onnx::MatMul_17765_11_0_173_12_40_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40" + ], + "const_args": [ + "onnx::MatMul_17766_11_0_174" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" + ], + "const_args": [ + "Attention_41_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" + ], + "const_args": [ + "Attention_41_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123" + ], + "const_args": [ + "Attention_41_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_41", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41" + ], + "const_args": [ + "onnx::MatMul_17777_11_0_175" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124" + ], + "const_args": [ + "onnx::MatMul_17778" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_41", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574" + ], + "const_args": [ + "onnx::MatMul_17788_11_0_176" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_41_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" + ], + "const_args": [ + "onnx::MatMul_17789_11_0_177_12_41_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_41_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125" + ], + "const_args": [ + "onnx::MatMul_17789_11_0_177_12_41_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41" + ], + "const_args": [ + "onnx::MatMul_17790_11_0_178" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" + ], + "const_args": [ + "Attention_42_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" + ], + "const_args": [ + "Attention_42_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126" + ], + "const_args": [ + "Attention_42_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_42", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42" + ], + "const_args": [ + "onnx::MatMul_17801_11_0_179" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127" + ], + "const_args": [ + "onnx::MatMul_17802" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_42", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582" + ], + "const_args": [ + "onnx::MatMul_17812_11_0_180" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_42_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" + ], + "const_args": [ + "onnx::MatMul_17813_11_0_181_12_42_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_42_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128" + ], + "const_args": [ + "onnx::MatMul_17813_11_0_181_12_42_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42" + ], + "const_args": [ + "onnx::MatMul_17814_11_0_182" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" + ], + "const_args": [ + "Attention_43_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" + ], + "const_args": [ + "Attention_43_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129" + ], + "const_args": [ + "Attention_43_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_43", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43" + ], + "const_args": [ + "onnx::MatMul_17825_11_0_183" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130" + ], + "const_args": [ + "onnx::MatMul_17826" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_SDCastBfp2Bf_/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_43", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_SDCastBf2Bfp_/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590" + ], + "const_args": [ + "onnx::MatMul_17836_11_0_184" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" + ], + "const_args": [ + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_43_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" + ], + "const_args": [ + "onnx::MatMul_17837_11_0_185_12_43_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_43_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131" + ], + "const_args": [ + "onnx::MatMul_17837_11_0_185_12_43_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43" + ], + "const_args": [ + "onnx::MatMul_17838_11_0_186" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155" + ], + "const_args": [ + "onnx::MatMul_17839_11_0_187" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594", + "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/Concat_1", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156", + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_24", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" + ], + "const_args": [ + "GroupNorm_24_wts_6_0_18" + ], + "out_args": [ + "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_24-/unet/up_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24" + ], + "const_args": [ + "NhwcConv_24_weight_NHWC" + ], + "out_args": [ + "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25", + "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_25", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157" + ], + "const_args": [ + "GroupNorm_25_wts_6_0_19" + ], + "out_args": [ + "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_25-/unet/up_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25" + ], + "const_args": [ + "NhwcConv_25_weight_NHWC" + ], + "out_args": [ + "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_26-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1" + ], + "const_args": [ + "NhwcConv_26_weight_NHWC" + ], + "out_args": [ + "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24", + "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_26", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" + ], + "const_args": [ + "GroupNorm_26_wts_6_1_6" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26" + ], + "const_args": [ + "onnx::MatMul_17846_11_0_188" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" + ], + "const_args": [ + "Attention_44_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" + ], + "const_args": [ + "Attention_44_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132" + ], + "const_args": [ + "Attention_44_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_44", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44" + ], + "const_args": [ + "onnx::MatMul_17857_11_0_189" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599", + "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133" + ], + "const_args": [ + "onnx::MatMul_17858" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_44", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600" + ], + "const_args": [ + "onnx::MatMul_17868_11_0_190" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_44_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" + ], + "const_args": [ + "onnx::MatMul_17869_11_0_191_12_44_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_44_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134" + ], + "const_args": [ + "onnx::MatMul_17869_11_0_191_12_44_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44" + ], + "const_args": [ + "onnx::MatMul_17870_11_0_192" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" + ], + "const_args": [ + "Attention_45_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" + ], + "const_args": [ + "Attention_45_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135" + ], + "const_args": [ + "Attention_45_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_45", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45" + ], + "const_args": [ + "onnx::MatMul_17881_11_0_193" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136" + ], + "const_args": [ + "onnx::MatMul_17882" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_45", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608" + ], + "const_args": [ + "onnx::MatMul_17892_11_0_194" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_45_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" + ], + "const_args": [ + "onnx::MatMul_17893_11_0_195_12_45_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_45_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137" + ], + "const_args": [ + "onnx::MatMul_17893_11_0_195_12_45_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45" + ], + "const_args": [ + "onnx::MatMul_17894_11_0_196" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" + ], + "const_args": [ + "Attention_46_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" + ], + "const_args": [ + "Attention_46_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138" + ], + "const_args": [ + "Attention_46_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_46", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46" + ], + "const_args": [ + "onnx::MatMul_17905_11_0_197" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139" + ], + "const_args": [ + "onnx::MatMul_17906" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_46", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616" + ], + "const_args": [ + "onnx::MatMul_17916_11_0_198" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_46_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" + ], + "const_args": [ + "onnx::MatMul_17917_11_0_199_12_46_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_46_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140" + ], + "const_args": [ + "onnx::MatMul_17917_11_0_199_12_46_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46" + ], + "const_args": [ + "onnx::MatMul_17918_11_0_200" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" + ], + "const_args": [ + "Attention_47_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" + ], + "const_args": [ + "Attention_47_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141" + ], + "const_args": [ + "Attention_47_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_47", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47" + ], + "const_args": [ + "onnx::MatMul_17929_11_0_201" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142" + ], + "const_args": [ + "onnx::MatMul_17930" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_47", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624" + ], + "const_args": [ + "onnx::MatMul_17940_11_0_202" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_47_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" + ], + "const_args": [ + "onnx::MatMul_17941_11_0_203_12_47_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_47_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143" + ], + "const_args": [ + "onnx::MatMul_17941_11_0_203_12_47_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47" + ], + "const_args": [ + "onnx::MatMul_17942_11_0_204" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" + ], + "const_args": [ + "Attention_48_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" + ], + "const_args": [ + "Attention_48_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144" + ], + "const_args": [ + "Attention_48_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_48", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48" + ], + "const_args": [ + "onnx::MatMul_17953_11_0_205" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145" + ], + "const_args": [ + "onnx::MatMul_17954" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_48", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632" + ], + "const_args": [ + "onnx::MatMul_17964_11_0_206" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_48_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" + ], + "const_args": [ + "onnx::MatMul_17965_11_0_207_12_48_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_48_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146" + ], + "const_args": [ + "onnx::MatMul_17965_11_0_207_12_48_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48" + ], + "const_args": [ + "onnx::MatMul_17966_11_0_208" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" + ], + "const_args": [ + "Attention_49_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" + ], + "const_args": [ + "Attention_49_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147" + ], + "const_args": [ + "Attention_49_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_49", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49" + ], + "const_args": [ + "onnx::MatMul_17977_11_0_209" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148" + ], + "const_args": [ + "onnx::MatMul_17978" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_49", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640" + ], + "const_args": [ + "onnx::MatMul_17988_11_0_210" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_49_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" + ], + "const_args": [ + "onnx::MatMul_17989_11_0_211_12_49_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_49_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149" + ], + "const_args": [ + "onnx::MatMul_17989_11_0_211_12_49_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49" + ], + "const_args": [ + "onnx::MatMul_17990_11_0_212" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" + ], + "const_args": [ + "Attention_50_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" + ], + "const_args": [ + "Attention_50_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150" + ], + "const_args": [ + "Attention_50_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_50", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50" + ], + "const_args": [ + "onnx::MatMul_18001_11_0_213" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151" + ], + "const_args": [ + "onnx::MatMul_18002" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_50", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648" + ], + "const_args": [ + "onnx::MatMul_18012_11_0_214" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_50_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" + ], + "const_args": [ + "onnx::MatMul_18013_11_0_215_12_50_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_50_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152" + ], + "const_args": [ + "onnx::MatMul_18013_11_0_215_12_50_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50" + ], + "const_args": [ + "onnx::MatMul_18014_11_0_216" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" + ], + "const_args": [ + "Attention_51_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" + ], + "const_args": [ + "Attention_51_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153" + ], + "const_args": [ + "Attention_51_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_51", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51" + ], + "const_args": [ + "onnx::MatMul_18025_11_0_217" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154" + ], + "const_args": [ + "onnx::MatMul_18026" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_51", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656" + ], + "const_args": [ + "onnx::MatMul_18036_11_0_218" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_51_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" + ], + "const_args": [ + "onnx::MatMul_18037_11_0_219_12_51_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_51_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155" + ], + "const_args": [ + "onnx::MatMul_18037_11_0_219_12_51_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51" + ], + "const_args": [ + "onnx::MatMul_18038_11_0_220" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" + ], + "const_args": [ + "Attention_52_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" + ], + "const_args": [ + "Attention_52_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156" + ], + "const_args": [ + "Attention_52_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_52", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52" + ], + "const_args": [ + "onnx::MatMul_18049_11_0_221" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157" + ], + "const_args": [ + "onnx::MatMul_18050" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_52", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664" + ], + "const_args": [ + "onnx::MatMul_18060_11_0_222" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_52_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" + ], + "const_args": [ + "onnx::MatMul_18061_11_0_223_12_52_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_52_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158" + ], + "const_args": [ + "onnx::MatMul_18061_11_0_223_12_52_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52" + ], + "const_args": [ + "onnx::MatMul_18062_11_0_224" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" + ], + "const_args": [ + "Attention_53_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" + ], + "const_args": [ + "Attention_53_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159" + ], + "const_args": [ + "Attention_53_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_53", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53" + ], + "const_args": [ + "onnx::MatMul_18073_11_0_225" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160" + ], + "const_args": [ + "onnx::MatMul_18074" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_SDCastBfp2Bf_/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_53", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_SDCastBf2Bfp_/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672" + ], + "const_args": [ + "onnx::MatMul_18084_11_0_226" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" + ], + "const_args": [ + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_53_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" + ], + "const_args": [ + "onnx::MatMul_18085_11_0_227_12_53_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_53_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161" + ], + "const_args": [ + "onnx::MatMul_18085_11_0_227_12_53_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53" + ], + "const_args": [ + "onnx::MatMul_18086_11_0_228" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188" + ], + "const_args": [ + "onnx::MatMul_18087_11_0_229" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676", + "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/Concat_2", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189", + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_27", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" + ], + "const_args": [ + "GroupNorm_27_wts_6_0_20" + ], + "out_args": [ + "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "3840" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_27-/unet/up_blocks.0/resnets.2/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27" + ], + "const_args": [ + "NhwcConv_27_weight_NHWC" + ], + "out_args": [ + "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28", + "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_28", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190" + ], + "const_args": [ + "GroupNorm_28_wts_6_0_21" + ], + "out_args": [ + "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_28-/unet/up_blocks.0/resnets.2/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28" + ], + "const_args": [ + "NhwcConv_28_weight_NHWC" + ], + "out_args": [ + "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_29-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2" + ], + "const_args": [ + "NhwcConv_29_weight_NHWC" + ], + "out_args": [ + "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/resnets.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27", + "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_29", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" + ], + "const_args": [ + "GroupNorm_29_wts_6_1_7" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29" + ], + "const_args": [ + "onnx::MatMul_18094_11_0_230" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" + ], + "const_args": [ + "Attention_54_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" + ], + "const_args": [ + "Attention_54_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162" + ], + "const_args": [ + "Attention_54_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_54", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54" + ], + "const_args": [ + "onnx::MatMul_18105_11_0_231" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681", + "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163" + ], + "const_args": [ + "onnx::MatMul_18106" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_54", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682" + ], + "const_args": [ + "onnx::MatMul_18116_11_0_232" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_54_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" + ], + "const_args": [ + "onnx::MatMul_18117_11_0_233_12_54_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_54_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164" + ], + "const_args": [ + "onnx::MatMul_18117_11_0_233_12_54_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54" + ], + "const_args": [ + "onnx::MatMul_18118_11_0_234" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" + ], + "const_args": [ + "Attention_55_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" + ], + "const_args": [ + "Attention_55_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165" + ], + "const_args": [ + "Attention_55_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_55", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55" + ], + "const_args": [ + "onnx::MatMul_18129_11_0_235" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166" + ], + "const_args": [ + "onnx::MatMul_18130" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_55", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690" + ], + "const_args": [ + "onnx::MatMul_18140_11_0_236" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_55_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" + ], + "const_args": [ + "onnx::MatMul_18141_11_0_237_12_55_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_55_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167" + ], + "const_args": [ + "onnx::MatMul_18141_11_0_237_12_55_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55" + ], + "const_args": [ + "onnx::MatMul_18142_11_0_238" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" + ], + "const_args": [ + "Attention_56_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" + ], + "const_args": [ + "Attention_56_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168" + ], + "const_args": [ + "Attention_56_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_56", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56" + ], + "const_args": [ + "onnx::MatMul_18153_11_0_239" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169" + ], + "const_args": [ + "onnx::MatMul_18154" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_56", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698" + ], + "const_args": [ + "onnx::MatMul_18164_11_0_240" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_56_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" + ], + "const_args": [ + "onnx::MatMul_18165_11_0_241_12_56_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/proj/MatMul/MatMulAddFusion_12_56_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170" + ], + "const_args": [ + "onnx::MatMul_18165_11_0_241_12_56_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56" + ], + "const_args": [ + "onnx::MatMul_18166_11_0_242" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" + ], + "const_args": [ + "Attention_57_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" + ], + "const_args": [ + "Attention_57_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171" + ], + "const_args": [ + "Attention_57_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_57", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57" + ], + "const_args": [ + "onnx::MatMul_18177_11_0_243" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172" + ], + "const_args": [ + "onnx::MatMul_18178" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_57", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706" + ], + "const_args": [ + "onnx::MatMul_18188_11_0_244" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_57_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" + ], + "const_args": [ + "onnx::MatMul_18189_11_0_245_12_57_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/proj/MatMul/MatMulAddFusion_12_57_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173" + ], + "const_args": [ + "onnx::MatMul_18189_11_0_245_12_57_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57" + ], + "const_args": [ + "onnx::MatMul_18190_11_0_246" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" + ], + "const_args": [ + "Attention_58_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" + ], + "const_args": [ + "Attention_58_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174" + ], + "const_args": [ + "Attention_58_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_58", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58" + ], + "const_args": [ + "onnx::MatMul_18201_11_0_247" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175" + ], + "const_args": [ + "onnx::MatMul_18202" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_58", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714" + ], + "const_args": [ + "onnx::MatMul_18212_11_0_248" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_58_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" + ], + "const_args": [ + "onnx::MatMul_18213_11_0_249_12_58_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/proj/MatMul/MatMulAddFusion_12_58_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176" + ], + "const_args": [ + "onnx::MatMul_18213_11_0_249_12_58_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58" + ], + "const_args": [ + "onnx::MatMul_18214_11_0_250" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" + ], + "const_args": [ + "Attention_59_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" + ], + "const_args": [ + "Attention_59_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177" + ], + "const_args": [ + "Attention_59_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_59", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59" + ], + "const_args": [ + "onnx::MatMul_18225_11_0_251" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178" + ], + "const_args": [ + "onnx::MatMul_18226" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_59", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722" + ], + "const_args": [ + "onnx::MatMul_18236_11_0_252" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_59_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" + ], + "const_args": [ + "onnx::MatMul_18237_11_0_253_12_59_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/proj/MatMul/MatMulAddFusion_12_59_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179" + ], + "const_args": [ + "onnx::MatMul_18237_11_0_253_12_59_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59" + ], + "const_args": [ + "onnx::MatMul_18238_11_0_254" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" + ], + "const_args": [ + "Attention_60_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" + ], + "const_args": [ + "Attention_60_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180" + ], + "const_args": [ + "Attention_60_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_60", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60" + ], + "const_args": [ + "onnx::MatMul_18249_11_0_255" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181" + ], + "const_args": [ + "onnx::MatMul_18250" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_60", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730" + ], + "const_args": [ + "onnx::MatMul_18260_11_0_256" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_60_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" + ], + "const_args": [ + "onnx::MatMul_18261_11_0_257_12_60_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/proj/MatMul/MatMulAddFusion_12_60_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182" + ], + "const_args": [ + "onnx::MatMul_18261_11_0_257_12_60_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60" + ], + "const_args": [ + "onnx::MatMul_18262_11_0_258" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" + ], + "const_args": [ + "Attention_61_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" + ], + "const_args": [ + "Attention_61_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183" + ], + "const_args": [ + "Attention_61_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_61", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61" + ], + "const_args": [ + "onnx::MatMul_18273_11_0_259" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184" + ], + "const_args": [ + "onnx::MatMul_18274" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_61", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738" + ], + "const_args": [ + "onnx::MatMul_18284_11_0_260" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_61_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" + ], + "const_args": [ + "onnx::MatMul_18285_11_0_261_12_61_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/proj/MatMul/MatMulAddFusion_12_61_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185" + ], + "const_args": [ + "onnx::MatMul_18285_11_0_261_12_61_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61" + ], + "const_args": [ + "onnx::MatMul_18286_11_0_262" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" + ], + "const_args": [ + "Attention_62_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" + ], + "const_args": [ + "Attention_62_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186" + ], + "const_args": [ + "Attention_62_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_62", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62" + ], + "const_args": [ + "onnx::MatMul_18297_11_0_263" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187" + ], + "const_args": [ + "onnx::MatMul_18298" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_62", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746" + ], + "const_args": [ + "onnx::MatMul_18308_11_0_264" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_62_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" + ], + "const_args": [ + "onnx::MatMul_18309_11_0_265_12_62_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/proj/MatMul/MatMulAddFusion_12_62_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188" + ], + "const_args": [ + "onnx::MatMul_18309_11_0_265_12_62_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62" + ], + "const_args": [ + "onnx::MatMul_18310_11_0_266" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" + ], + "const_args": [ + "Attention_63_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" + ], + "const_args": [ + "Attention_63_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189" + ], + "const_args": [ + "Attention_63_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_63", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "1024" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63" + ], + "const_args": [ + "onnx::MatMul_18321_11_0_267" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190" + ], + "const_args": [ + "onnx::MatMul_18322" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "20" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_SDCastBfp2Bf_/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "20", + "1024", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_63", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "20" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_SDCastBf2Bfp_/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127" + ], + "const_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754" + ], + "const_args": [ + "onnx::MatMul_18332_11_0_268" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" + ], + "const_args": [ + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_63_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" + ], + "const_args": [ + "onnx::MatMul_18333_11_0_269_12_63_0" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/proj/MatMul/MatMulAddFusion_12_63_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191" + ], + "const_args": [ + "onnx::MatMul_18333_11_0_269_12_63_1" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63" + ], + "const_args": [ + "onnx::MatMul_18334_11_0_270" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "5120" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221" + ], + "const_args": [ + "onnx::MatMul_18335_11_0_271" + ], + "out_args": [ + "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/attentions.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758", + "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.0/upsamplers.0/Resize", + "type": "SDResize_bfp", + "in_args": [ + "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222" + ], + "const_args": [ + "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0" + ], + "out_args": [ + "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_30-/unet/up_blocks.0/upsamplers.0/conv/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0" + ], + "const_args": [ + "NhwcConv_30_weight_NHWC" + ], + "out_args": [ + "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/Concat", + "type": "SDConcat_bfp", + "in_args": [ + "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30", + "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_30", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" + ], + "const_args": [ + "GroupNorm_30_wts_6_0_22" + ], + "out_args": [ + "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1920" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "3840" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_31-/unet/up_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30" + ], + "const_args": [ + "NhwcConv_31_weight_NHWC" + ], + "out_args": [ + "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32", + "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_31", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223" + ], + "const_args": [ + "GroupNorm_31_wts_6_0_23" + ], + "out_args": [ + "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_32-/unet/up_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31" + ], + "const_args": [ + "NhwcConv_32_weight_NHWC" + ], + "out_args": [ + "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_33-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3" + ], + "const_args": [ + "NhwcConv_33_weight_NHWC" + ], + "out_args": [ + "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31", + "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_32", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" + ], + "const_args": [ + "GroupNorm_32_wts_6_1_8" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32" + ], + "const_args": [ + "onnx::MatMul_18343_11_0_272" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" + ], + "const_args": [ + "Attention_64_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" + ], + "const_args": [ + "Attention_64_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192" + ], + "const_args": [ + "Attention_64_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_64", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64" + ], + "const_args": [ + "onnx::MatMul_18354_11_0_273" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763", + "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193" + ], + "const_args": [ + "onnx::MatMul_18355" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_SDCastBfp2Bf_/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_64", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_SDCastBf2Bfp_/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764" + ], + "const_args": [ + "onnx::MatMul_18365_11_0_274" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_64_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" + ], + "const_args": [ + "onnx::MatMul_18366_11_0_275_12_64_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_64_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194" + ], + "const_args": [ + "onnx::MatMul_18366_11_0_275_12_64_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64" + ], + "const_args": [ + "onnx::MatMul_18367_11_0_276" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" + ], + "const_args": [ + "Attention_65_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" + ], + "const_args": [ + "Attention_65_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195" + ], + "const_args": [ + "Attention_65_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_65", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65" + ], + "const_args": [ + "onnx::MatMul_18378_11_0_277" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196" + ], + "const_args": [ + "onnx::MatMul_18379" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_SDCastBfp2Bf_/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_65", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_SDCastBf2Bfp_/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772" + ], + "const_args": [ + "onnx::MatMul_18389_11_0_278" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" + ], + "const_args": [ + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_65_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" + ], + "const_args": [ + "onnx::MatMul_18390_11_0_279_12_65_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_65_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197" + ], + "const_args": [ + "onnx::MatMul_18390_11_0_279_12_65_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65" + ], + "const_args": [ + "onnx::MatMul_18391_11_0_280" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230" + ], + "const_args": [ + "onnx::MatMul_18392_11_0_281" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776", + "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/Concat_1", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231", + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_33", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" + ], + "const_args": [ + "GroupNorm_33_wts_6_0_24" + ], + "out_args": [ + "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_34-/unet/up_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33" + ], + "const_args": [ + "NhwcConv_34_weight_NHWC" + ], + "out_args": [ + "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35", + "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_34", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232" + ], + "const_args": [ + "GroupNorm_34_wts_6_0_25" + ], + "out_args": [ + "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_35-/unet/up_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34" + ], + "const_args": [ + "NhwcConv_35_weight_NHWC" + ], + "out_args": [ + "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_36-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4" + ], + "const_args": [ + "NhwcConv_36_weight_NHWC" + ], + "out_args": [ + "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34", + "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_35", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" + ], + "const_args": [ + "GroupNorm_35_wts_6_1_9" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35" + ], + "const_args": [ + "onnx::MatMul_18399_11_0_282" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" + ], + "const_args": [ + "Attention_66_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" + ], + "const_args": [ + "Attention_66_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198" + ], + "const_args": [ + "Attention_66_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_66", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66" + ], + "const_args": [ + "onnx::MatMul_18410_11_0_283" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781", + "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199" + ], + "const_args": [ + "onnx::MatMul_18411" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_SDCastBfp2Bf_/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_66", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_SDCastBf2Bfp_/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782" + ], + "const_args": [ + "onnx::MatMul_18421_11_0_284" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_66_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" + ], + "const_args": [ + "onnx::MatMul_18422_11_0_285_12_66_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_66_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200" + ], + "const_args": [ + "onnx::MatMul_18422_11_0_285_12_66_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66" + ], + "const_args": [ + "onnx::MatMul_18423_11_0_286" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" + ], + "const_args": [ + "Attention_67_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" + ], + "const_args": [ + "Attention_67_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201" + ], + "const_args": [ + "Attention_67_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_67", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67" + ], + "const_args": [ + "onnx::MatMul_18434_11_0_287" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202" + ], + "const_args": [ + "onnx::MatMul_18435" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_SDCastBfp2Bf_/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_67", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_SDCastBf2Bfp_/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790" + ], + "const_args": [ + "onnx::MatMul_18445_11_0_288" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" + ], + "const_args": [ + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_67_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" + ], + "const_args": [ + "onnx::MatMul_18446_11_0_289_12_67_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_67_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203" + ], + "const_args": [ + "onnx::MatMul_18446_11_0_289_12_67_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67" + ], + "const_args": [ + "onnx::MatMul_18447_11_0_290" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239" + ], + "const_args": [ + "onnx::MatMul_18448_11_0_291" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794", + "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/Concat_2", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240", + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_36", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" + ], + "const_args": [ + "GroupNorm_36_wts_6_0_26" + ], + "out_args": [ + "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "960" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "960" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_37-/unet/up_blocks.1/resnets.2/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36" + ], + "const_args": [ + "NhwcConv_37_weight_NHWC" + ], + "out_args": [ + "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38", + "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_37", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241" + ], + "const_args": [ + "GroupNorm_37_wts_6_0_27" + ], + "out_args": [ + "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_38-/unet/up_blocks.1/resnets.2/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37" + ], + "const_args": [ + "NhwcConv_38_weight_NHWC" + ], + "out_args": [ + "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_39-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5" + ], + "const_args": [ + "NhwcConv_39_weight_NHWC" + ], + "out_args": [ + "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/resnets.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37", + "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_38", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" + ], + "const_args": [ + "GroupNorm_38_wts_6_1_10" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/proj_in/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38" + ], + "const_args": [ + "onnx::MatMul_18455_11_0_292" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" + ], + "const_args": [ + "Attention_68_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" + ], + "const_args": [ + "Attention_68_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204" + ], + "const_args": [ + "Attention_68_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_68", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68" + ], + "const_args": [ + "onnx::MatMul_18466_11_0_293" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799", + "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205" + ], + "const_args": [ + "onnx::MatMul_18467" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_SDCastBfp2Bf_/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_68", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_SDCastBf2Bfp_/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800" + ], + "const_args": [ + "onnx::MatMul_18477_11_0_294" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_68_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" + ], + "const_args": [ + "onnx::MatMul_18478_11_0_295_12_68_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul/MatMulAddFusion_12_68_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206" + ], + "const_args": [ + "onnx::MatMul_18478_11_0_295_12_68_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68" + ], + "const_args": [ + "onnx::MatMul_18479_11_0_296" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_key", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" + ], + "const_args": [ + "Attention_69_qkv_weight_key" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_query", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" + ], + "const_args": [ + "Attention_69_qkv_weight_query" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_MatMul_value", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207" + ], + "const_args": [ + "Attention_69_qkv_weight_value" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Attention_69", + "type": "SDMHA_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "4096" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69" + ], + "const_args": [ + "onnx::MatMul_18490_11_0_297" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208" + ], + "const_args": [ + "onnx::MatMul_18491" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "10" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_SDCastBfp2Bf_/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", + "type": "SDCastBfp2Bf", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "10", + "4096", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_69", + "type": "SDFlatMHA", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "10" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "64", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.1" + ] + }, + "is_flat_mha_1_1": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_SDCastBf2Bfp_/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808", + "type": "SDCastBf2Bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139" + ], + "const_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808" + ], + "const_args": [ + "onnx::MatMul_18501_11_0_298" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization", + "type": "SDLayerNorm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" + ], + "const_args": [ + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_69_0", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" + ], + "const_args": [ + "onnx::MatMul_18502_11_0_299_12_69_0" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/proj/MatMul/MatMulAddFusion_12_69_1", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209" + ], + "const_args": [ + "onnx::MatMul_18502_11_0_299_12_69_1" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4", + "type": "SDMul_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69" + ], + "const_args": [ + "onnx::MatMul_18503_11_0_300" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/proj_out/MatMul/MatMulAddFusion", + "type": "SDGemm_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248" + ], + "const_args": [ + "onnx::MatMul_18504_11_0_301" + ], + "out_args": [ + "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/attentions.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812", + "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.1/upsamplers.0/Resize", + "type": "SDResize_bfp", + "in_args": [ + "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249" + ], + "const_args": [ + "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1" + ], + "out_args": [ + "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_40-/unet/up_blocks.1/upsamplers.0/conv/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1" + ], + "const_args": [ + "NhwcConv_40_weight_NHWC" + ], + "out_args": [ + "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/Concat", + "type": "SDConcat_bfp", + "in_args": [ + "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40", + "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_39", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" + ], + "const_args": [ + "GroupNorm_39_wts_6_0_28" + ], + "out_args": [ + "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "960" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "960" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "960" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_41-/unet/up_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39" + ], + "const_args": [ + "NhwcConv_41_weight_NHWC" + ], + "out_args": [ + "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.0/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42", + "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_40", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250" + ], + "const_args": [ + "GroupNorm_40_wts_6_0_29" + ], + "out_args": [ + "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_42-/unet/up_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40" + ], + "const_args": [ + "NhwcConv_42_weight_NHWC" + ], + "out_args": [ + "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_43-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6" + ], + "const_args": [ + "NhwcConv_43_weight_NHWC" + ], + "out_args": [ + "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.0/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41", + "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/Concat_1", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251", + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_41", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" + ], + "const_args": [ + "GroupNorm_41_wts_6_0_30" + ], + "out_args": [ + "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_44-/unet/up_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41" + ], + "const_args": [ + "NhwcConv_44_weight_NHWC" + ], + "out_args": [ + "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.1/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45", + "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_42", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252" + ], + "const_args": [ + "GroupNorm_42_wts_6_0_31" + ], + "out_args": [ + "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_45-/unet/up_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42" + ], + "const_args": [ + "NhwcConv_45_weight_NHWC" + ], + "out_args": [ + "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_46-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7" + ], + "const_args": [ + "NhwcConv_46_weight_NHWC" + ], + "out_args": [ + "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.1/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44", + "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/Concat_2", + "type": "SDConcat_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253", + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_43", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" + ], + "const_args": [ + "GroupNorm_43_wts_6_0_32" + ], + "out_args": [ + "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_47-/unet/up_blocks.2/resnets.2/conv1/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43" + ], + "const_args": [ + "NhwcConv_47_weight_NHWC" + ], + "out_args": [ + "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.2/Add", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48", + "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_44", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254" + ], + "const_args": [ + "GroupNorm_44_wts_6_0_33" + ], + "out_args": [ + "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_48-/unet/up_blocks.2/resnets.2/conv2/Conv", + "type": "SDConv_bfp", + "in_args": [ + "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44" + ], + "const_args": [ + "NhwcConv_48_weight_NHWC" + ], + "out_args": [ + "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_49-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv", + "type": "SDConv_bfp", + "in_args": [ + "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8" + ], + "const_args": [ + "NhwcConv_49_weight_NHWC" + ], + "out_args": [ + "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/unet/up_blocks.2/resnets.2/Add_1", + "type": "SDAdd_bfp", + "in_args": [ + "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47", + "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49" + ], + "const_args": [], + "out_args": [ + "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_45", + "type": "SDGroupNorm_bfp", + "in_args": [ + "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255" + ], + "const_args": [ + "GroupNorm_45_wts_6_0_34" + ], + "out_args": [ + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_SDCastBfp2Bf_GroupNorm_45.out7_45.out6_0_34", + "type": "SDCastBfp2Bf", + "in_args": [ + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45" + ], + "const_args": [ + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts" + ], + "out_args": [ + "GroupNorm_45.out7_45.out6_0_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_50-/unet/conv_out/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_45.out7_45.out6_0_34" + ], + "const_args": [ + "NhwcConv_50_weight_NHWC" + ], + "out_args": [ + "NhwcConv_50_out-out_sample.out0_1_50" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "128", + "128", + "4" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "4", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 88584448, + "xrt_arg_id": 0, + "packed_tensors": [ + "Transpose_34_out-sample.out0_1_0", + "/unet/Concat_3_output_0.out2_3_1", + "encoder_hidden_states.out2_6_0", + "/unet/time_proj/Concat_1_output_0.out2_3_0" + ] + }, + "out": { + "buffer_size": 262144, + "xrt_arg_id": 1, + "packed_tensors": [ + "NhwcConv_50_out-out_sample.out0_1_50" + ] + }, + "scratch": { + "buffer_size": 7679737600, + "xrt_arg_id": 2, + "packed_tensors": [ + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0", + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0", + "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0", + "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1", + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1", + "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67", + "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0", + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0", + "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2", + "/unet/Add_output_0.out_17_1_0", + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2", + "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0", + "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1", + "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1", + "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1", + "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2", + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2", + "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2", + "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1", + "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3", + "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3", + "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3", + "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4", + "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4", + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5", + "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4", + "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2", + "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7", + "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5", + "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5", + "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3", + "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4", + "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5", + "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6", + "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127", + "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8", + "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9", + "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139", + "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11", + "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12", + "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13", + "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14", + "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15", + "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16", + "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8", + "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6", + "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6", + "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6", + "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12", + "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248", + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13", + "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7", + "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9", + "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14", + "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8", + "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10", + "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15", + "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9", + "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21", + "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266", + "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22", + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11", + "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10", + "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13", + "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23", + "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11", + "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14", + "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12", + "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24", + "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12", + "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54", + "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348", + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55", + "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13", + "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15", + "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56", + "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14", + "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16", + "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57", + "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15", + "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87", + "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430", + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88", + "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16", + "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17", + "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89", + "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17", + "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18", + "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90", + "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18", + "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92", + "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74", + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437", + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438", + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24", + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439", + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95", + "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77", + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445", + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446", + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25", + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447", + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98", + "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80", + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453", + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454", + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26", + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455", + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101", + "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83", + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461", + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462", + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27", + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463", + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104", + "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86", + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469", + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470", + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28", + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471", + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107", + "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89", + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477", + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478", + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29", + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479", + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110", + "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92", + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485", + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486", + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30", + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487", + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113", + "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95", + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493", + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494", + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31", + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495", + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116", + "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98", + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501", + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502", + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32", + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503", + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507", + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508", + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119", + "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101", + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509", + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510", + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33", + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511", + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120", + "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512", + "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121", + "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19", + "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19", + "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122", + "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20", + "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20", + "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123", + "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0", + "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21", + "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22", + "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124", + "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22", + "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23", + "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21", + "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125", + "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23", + "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155", + "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594", + "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156", + "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1", + "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24", + "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25", + "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157", + "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25", + "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26", + "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24", + "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158", + "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26", + "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188", + "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676", + "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189", + "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2", + "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27", + "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28", + "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190", + "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28", + "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29", + "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27", + "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191", + "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29", + "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221", + "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758", + "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222", + "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0", + "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30", + "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3", + "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30", + "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32", + "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223", + "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31", + "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33", + "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31", + "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224", + "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32", + "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230", + "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776", + "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231", + "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4", + "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33", + "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35", + "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232", + "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34", + "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36", + "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34", + "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233", + "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35", + "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239", + "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794", + "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240", + "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5", + "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36", + "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38", + "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241", + "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37", + "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39", + "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37", + "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242", + "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38", + "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248", + "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812", + "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249", + "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1", + "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40", + "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6", + "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39", + "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42", + "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250", + "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40", + "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43", + "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41", + "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251", + "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7", + "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41", + "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45", + "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252", + "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42", + "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46", + "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44", + "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253", + "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8", + "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43", + "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48", + "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254", + "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44", + "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49", + "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47", + "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255", + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45", + "GroupNorm_45.out7_45.out6_0_34" + ] + }, + "const": { + "buffer_size": 2936555456, + "xrt_arg_id": 3, + "packed_tensors": [ + "NhwcConv_0_weight_NHWC", + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts", + "GroupNorm_0_wts_6_0_0", + "unet.add_embedding.linear_1.weight_11_1_0", + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1", + "unet.add_embedding.linear_2.weight_11_1_1", + "onnx::MatMul_16751", + "onnx::MatMul_16752", + "onnx::MatMul_16775", + "onnx::MatMul_16776", + "onnx::MatMul_16807", + "onnx::MatMul_16808", + "onnx::MatMul_16831", + "onnx::MatMul_16832", + "onnx::MatMul_16863", + "onnx::MatMul_16864", + "onnx::MatMul_16887", + "onnx::MatMul_16888", + "onnx::MatMul_16911", + "onnx::MatMul_16912", + "onnx::MatMul_16935", + "onnx::MatMul_16936", + "onnx::MatMul_16959", + "onnx::MatMul_16960", + "onnx::MatMul_16983", + "onnx::MatMul_16984", + "onnx::MatMul_17007", + "onnx::MatMul_17008", + "onnx::MatMul_17031", + "onnx::MatMul_17032", + "onnx::MatMul_17055", + "onnx::MatMul_17056", + "onnx::MatMul_17079", + "onnx::MatMul_17080", + "onnx::MatMul_17111", + "onnx::MatMul_17112", + "onnx::MatMul_17135", + "onnx::MatMul_17136", + "onnx::MatMul_17159", + "onnx::MatMul_17160", + "onnx::MatMul_17183", + "onnx::MatMul_17184", + "onnx::MatMul_17207", + "onnx::MatMul_17208", + "onnx::MatMul_17231", + "onnx::MatMul_17232", + "onnx::MatMul_17255", + "onnx::MatMul_17256", + "onnx::MatMul_17279", + "onnx::MatMul_17280", + "onnx::MatMul_17303", + "onnx::MatMul_17304", + "onnx::MatMul_17327", + "onnx::MatMul_17328", + "onnx::MatMul_17359", + "onnx::MatMul_17360", + "onnx::MatMul_17383", + "onnx::MatMul_17384", + "onnx::MatMul_17407", + "onnx::MatMul_17408", + "onnx::MatMul_17431", + "onnx::MatMul_17432", + "onnx::MatMul_17455", + "onnx::MatMul_17456", + "onnx::MatMul_17479", + "onnx::MatMul_17480", + "onnx::MatMul_17503", + "onnx::MatMul_17504", + "onnx::MatMul_17527", + "onnx::MatMul_17528", + "onnx::MatMul_17551", + "onnx::MatMul_17552", + "onnx::MatMul_17575", + "onnx::MatMul_17576", + "unet.time_embedding.linear_1.weight_11_1_2", + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0", + "unet.time_embedding.linear_2.weight_11_1_3", + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2", + "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4", + "NhwcConv_1_weight_NHWC", + "GroupNorm_1_wts_6_0_1", + "NhwcConv_2_weight_NHWC", + "GroupNorm_2_wts_6_0_2", + "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5", + "NhwcConv_3_weight_NHWC", + "GroupNorm_3_wts_6_0_3", + "NhwcConv_4_weight_NHWC", + "NhwcConv_5_weight_NHWC", + "GroupNorm_4_wts_6_0_4", + "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6", + "NhwcConv_6_weight_NHWC", + "GroupNorm_5_wts_6_0_5", + "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7", + "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8", + "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9", + "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10", + "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11", + "onnx::MatMul_17611", + "onnx::MatMul_17612", + "onnx::MatMul_17635", + "onnx::MatMul_17636", + "onnx::MatMul_17659", + "onnx::MatMul_17660", + "onnx::MatMul_17683", + "onnx::MatMul_17684", + "onnx::MatMul_17707", + "onnx::MatMul_17708", + "onnx::MatMul_17731", + "onnx::MatMul_17732", + "onnx::MatMul_17755", + "onnx::MatMul_17756", + "onnx::MatMul_17779", + "onnx::MatMul_17780", + "onnx::MatMul_17803", + "onnx::MatMul_17804", + "onnx::MatMul_17827", + "onnx::MatMul_17828", + "onnx::MatMul_17859", + "onnx::MatMul_17860", + "onnx::MatMul_17883", + "onnx::MatMul_17884", + "onnx::MatMul_17907", + "onnx::MatMul_17908", + "onnx::MatMul_17931", + "onnx::MatMul_17932", + "onnx::MatMul_17955", + "onnx::MatMul_17956", + "onnx::MatMul_17979", + "onnx::MatMul_17980", + "onnx::MatMul_18003", + "onnx::MatMul_18004", + "onnx::MatMul_18027", + "onnx::MatMul_18028", + "onnx::MatMul_18051", + "onnx::MatMul_18052", + "onnx::MatMul_18075", + "onnx::MatMul_18076", + "onnx::MatMul_18107", + "onnx::MatMul_18108", + "onnx::MatMul_18131", + "onnx::MatMul_18132", + "onnx::MatMul_18155", + "onnx::MatMul_18156", + "onnx::MatMul_18179", + "onnx::MatMul_18180", + "onnx::MatMul_18203", + "onnx::MatMul_18204", + "onnx::MatMul_18227", + "onnx::MatMul_18228", + "onnx::MatMul_18251", + "onnx::MatMul_18252", + "onnx::MatMul_18275", + "onnx::MatMul_18276", + "onnx::MatMul_18299", + "onnx::MatMul_18300", + "onnx::MatMul_18323", + "onnx::MatMul_18324", + "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12", + "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13", + "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14", + "onnx::MatMul_18356", + "onnx::MatMul_18357", + "onnx::MatMul_18380", + "onnx::MatMul_18381", + "onnx::MatMul_18412", + "onnx::MatMul_18413", + "onnx::MatMul_18436", + "onnx::MatMul_18437", + "onnx::MatMul_18468", + "onnx::MatMul_18469", + "onnx::MatMul_18492", + "onnx::MatMul_18493", + "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15", + "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16", + "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17", + "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18", + "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19", + "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20", + "NhwcConv_7_weight_NHWC", + "NhwcConv_8_weight_NHWC", + "GroupNorm_6_wts_6_1_0", + "onnx::MatMul_16738_11_0_0", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", + "Attention_0_qkv_weight_key", + "Attention_0_qkv_weight_query", + "Attention_0_qkv_weight_value", + "onnx::MatMul_16749_11_0_1", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_16750", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts", + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts", + "onnx::MatMul_16760_11_0_2", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_16761_11_0_3_12_0_0", + "onnx::MatMul_16761_11_0_3_12_0_1", + "onnx::MatMul_16762_11_0_4", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias", + "Attention_1_qkv_weight_key", + "Attention_1_qkv_weight_query", + "Attention_1_qkv_weight_value", + "onnx::MatMul_16773_11_0_5", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias", + "onnx::MatMul_16774", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts", + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts", + "onnx::MatMul_16784_11_0_6", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias", + "onnx::MatMul_16785_11_0_7_12_1_0", + "onnx::MatMul_16785_11_0_7_12_1_1", + "onnx::MatMul_16786_11_0_8", + "onnx::MatMul_16787_11_0_9", + "GroupNorm_7_wts_6_0_6", + "NhwcConv_9_weight_NHWC", + "GroupNorm_8_wts_6_0_7", + "NhwcConv_10_weight_NHWC", + "GroupNorm_9_wts_6_1_1", + "onnx::MatMul_16794_11_0_10", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", + "Attention_2_qkv_weight_key", + "Attention_2_qkv_weight_query", + "Attention_2_qkv_weight_value", + "onnx::MatMul_16805_11_0_11", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_16806", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts", + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts", + "onnx::MatMul_16816_11_0_12", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_16817_11_0_13_12_2_0", + "onnx::MatMul_16817_11_0_13_12_2_1", + "onnx::MatMul_16818_11_0_14", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias", + "Attention_3_qkv_weight_key", + "Attention_3_qkv_weight_query", + "Attention_3_qkv_weight_value", + "onnx::MatMul_16829_11_0_15", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias", + "onnx::MatMul_16830", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts", + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts", + "onnx::MatMul_16840_11_0_16", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias", + "onnx::MatMul_16841_11_0_17_12_3_0", + "onnx::MatMul_16841_11_0_17_12_3_1", + "onnx::MatMul_16842_11_0_18", + "onnx::MatMul_16843_11_0_19", + "NhwcConv_11_weight_NHWC", + "GroupNorm_10_wts_6_0_8", + "NhwcConv_12_weight_NHWC", + "GroupNorm_11_wts_6_0_9", + "NhwcConv_13_weight_NHWC", + "NhwcConv_14_weight_NHWC", + "GroupNorm_12_wts_6_1_2", + "onnx::MatMul_16850_11_0_20", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias", + "Attention_4_qkv_weight_key", + "Attention_4_qkv_weight_query", + "Attention_4_qkv_weight_value", + "onnx::MatMul_16861_11_0_21", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_16862", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts", + "onnx::MatMul_16872_11_0_22", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_16873_11_0_23_12_4_0", + "onnx::MatMul_16873_11_0_23_12_4_1", + "onnx::MatMul_16874_11_0_24", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias", + "Attention_5_qkv_weight_key", + "Attention_5_qkv_weight_query", + "Attention_5_qkv_weight_value", + "onnx::MatMul_16885_11_0_25", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias", + "onnx::MatMul_16886", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts", + "onnx::MatMul_16896_11_0_26", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias", + "onnx::MatMul_16897_11_0_27_12_5_0", + "onnx::MatMul_16897_11_0_27_12_5_1", + "onnx::MatMul_16898_11_0_28", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias", + "Attention_6_qkv_weight_key", + "Attention_6_qkv_weight_query", + "Attention_6_qkv_weight_value", + "onnx::MatMul_16909_11_0_29", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias", + "onnx::MatMul_16910", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts", + "onnx::MatMul_16920_11_0_30", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias", + "onnx::MatMul_16921_11_0_31_12_6_0", + "onnx::MatMul_16921_11_0_31_12_6_1", + "onnx::MatMul_16922_11_0_32", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias", + "Attention_7_qkv_weight_key", + "Attention_7_qkv_weight_query", + "Attention_7_qkv_weight_value", + "onnx::MatMul_16933_11_0_33", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias", + "onnx::MatMul_16934", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts", + "onnx::MatMul_16944_11_0_34", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias", + "onnx::MatMul_16945_11_0_35_12_7_0", + "onnx::MatMul_16945_11_0_35_12_7_1", + "onnx::MatMul_16946_11_0_36", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias", + "Attention_8_qkv_weight_key", + "Attention_8_qkv_weight_query", + "Attention_8_qkv_weight_value", + "onnx::MatMul_16957_11_0_37", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias", + "onnx::MatMul_16958", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts", + "onnx::MatMul_16968_11_0_38", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias", + "onnx::MatMul_16969_11_0_39_12_8_0", + "onnx::MatMul_16969_11_0_39_12_8_1", + "onnx::MatMul_16970_11_0_40", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias", + "Attention_9_qkv_weight_key", + "Attention_9_qkv_weight_query", + "Attention_9_qkv_weight_value", + "onnx::MatMul_16981_11_0_41", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias", + "onnx::MatMul_16982", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts", + "onnx::MatMul_16992_11_0_42", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias", + "onnx::MatMul_16993_11_0_43_12_9_0", + "onnx::MatMul_16993_11_0_43_12_9_1", + "onnx::MatMul_16994_11_0_44", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias", + "Attention_10_qkv_weight_key", + "Attention_10_qkv_weight_query", + "Attention_10_qkv_weight_value", + "onnx::MatMul_17005_11_0_45", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias", + "onnx::MatMul_17006", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts", + "onnx::MatMul_17016_11_0_46", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias", + "onnx::MatMul_17017_11_0_47_12_10_0", + "onnx::MatMul_17017_11_0_47_12_10_1", + "onnx::MatMul_17018_11_0_48", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias", + "Attention_11_qkv_weight_key", + "Attention_11_qkv_weight_query", + "Attention_11_qkv_weight_value", + "onnx::MatMul_17029_11_0_49", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias", + "onnx::MatMul_17030", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts", + "onnx::MatMul_17040_11_0_50", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias", + "onnx::MatMul_17041_11_0_51_12_11_0", + "onnx::MatMul_17041_11_0_51_12_11_1", + "onnx::MatMul_17042_11_0_52", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias", + "Attention_12_qkv_weight_key", + "Attention_12_qkv_weight_query", + "Attention_12_qkv_weight_value", + "onnx::MatMul_17053_11_0_53", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias", + "onnx::MatMul_17054", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts", + "onnx::MatMul_17064_11_0_54", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias", + "onnx::MatMul_17065_11_0_55_12_12_0", + "onnx::MatMul_17065_11_0_55_12_12_1", + "onnx::MatMul_17066_11_0_56", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias", + "Attention_13_qkv_weight_key", + "Attention_13_qkv_weight_query", + "Attention_13_qkv_weight_value", + "onnx::MatMul_17077_11_0_57", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias", + "onnx::MatMul_17078", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts", + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts", + "onnx::MatMul_17088_11_0_58", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight", + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias", + "onnx::MatMul_17089_11_0_59_12_13_0", + "onnx::MatMul_17089_11_0_59_12_13_1", + "onnx::MatMul_17090_11_0_60", + "onnx::MatMul_17091_11_0_61", + "GroupNorm_13_wts_6_0_10", + "NhwcConv_15_weight_NHWC", + "GroupNorm_14_wts_6_0_11", + "NhwcConv_16_weight_NHWC", + "GroupNorm_15_wts_6_1_3", + "onnx::MatMul_17098_11_0_62", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias", + "Attention_14_qkv_weight_key", + "Attention_14_qkv_weight_query", + "Attention_14_qkv_weight_value", + "onnx::MatMul_17109_11_0_63", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_17110", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts", + "onnx::MatMul_17120_11_0_64", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_17121_11_0_65_12_14_0", + "onnx::MatMul_17121_11_0_65_12_14_1", + "onnx::MatMul_17122_11_0_66", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias", + "Attention_15_qkv_weight_key", + "Attention_15_qkv_weight_query", + "Attention_15_qkv_weight_value", + "onnx::MatMul_17133_11_0_67", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias", + "onnx::MatMul_17134", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts", + "onnx::MatMul_17144_11_0_68", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias", + "onnx::MatMul_17145_11_0_69_12_15_0", + "onnx::MatMul_17145_11_0_69_12_15_1", + "onnx::MatMul_17146_11_0_70", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias", + "Attention_16_qkv_weight_key", + "Attention_16_qkv_weight_query", + "Attention_16_qkv_weight_value", + "onnx::MatMul_17157_11_0_71", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias", + "onnx::MatMul_17158", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts", + "onnx::MatMul_17168_11_0_72", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias", + "onnx::MatMul_17169_11_0_73_12_16_0", + "onnx::MatMul_17169_11_0_73_12_16_1", + "onnx::MatMul_17170_11_0_74", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias", + "Attention_17_qkv_weight_key", + "Attention_17_qkv_weight_query", + "Attention_17_qkv_weight_value", + "onnx::MatMul_17181_11_0_75", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias", + "onnx::MatMul_17182", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts", + "onnx::MatMul_17192_11_0_76", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias", + "onnx::MatMul_17193_11_0_77_12_17_0", + "onnx::MatMul_17193_11_0_77_12_17_1", + "onnx::MatMul_17194_11_0_78", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias", + "Attention_18_qkv_weight_key", + "Attention_18_qkv_weight_query", + "Attention_18_qkv_weight_value", + "onnx::MatMul_17205_11_0_79", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias", + "onnx::MatMul_17206", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts", + "onnx::MatMul_17216_11_0_80", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias", + "onnx::MatMul_17217_11_0_81_12_18_0", + "onnx::MatMul_17217_11_0_81_12_18_1", + "onnx::MatMul_17218_11_0_82", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias", + "Attention_19_qkv_weight_key", + "Attention_19_qkv_weight_query", + "Attention_19_qkv_weight_value", + "onnx::MatMul_17229_11_0_83", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias", + "onnx::MatMul_17230", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts", + "onnx::MatMul_17240_11_0_84", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias", + "onnx::MatMul_17241_11_0_85_12_19_0", + "onnx::MatMul_17241_11_0_85_12_19_1", + "onnx::MatMul_17242_11_0_86", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias", + "Attention_20_qkv_weight_key", + "Attention_20_qkv_weight_query", + "Attention_20_qkv_weight_value", + "onnx::MatMul_17253_11_0_87", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias", + "onnx::MatMul_17254", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts", + "onnx::MatMul_17264_11_0_88", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias", + "onnx::MatMul_17265_11_0_89_12_20_0", + "onnx::MatMul_17265_11_0_89_12_20_1", + "onnx::MatMul_17266_11_0_90", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias", + "Attention_21_qkv_weight_key", + "Attention_21_qkv_weight_query", + "Attention_21_qkv_weight_value", + "onnx::MatMul_17277_11_0_91", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias", + "onnx::MatMul_17278", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts", + "onnx::MatMul_17288_11_0_92", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias", + "onnx::MatMul_17289_11_0_93_12_21_0", + "onnx::MatMul_17289_11_0_93_12_21_1", + "onnx::MatMul_17290_11_0_94", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias", + "Attention_22_qkv_weight_key", + "Attention_22_qkv_weight_query", + "Attention_22_qkv_weight_value", + "onnx::MatMul_17301_11_0_95", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias", + "onnx::MatMul_17302", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts", + "onnx::MatMul_17312_11_0_96", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias", + "onnx::MatMul_17313_11_0_97_12_22_0", + "onnx::MatMul_17313_11_0_97_12_22_1", + "onnx::MatMul_17314_11_0_98", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias", + "Attention_23_qkv_weight_key", + "Attention_23_qkv_weight_query", + "Attention_23_qkv_weight_value", + "onnx::MatMul_17325_11_0_99", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias", + "onnx::MatMul_17326", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts", + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts", + "onnx::MatMul_17336_11_0_100", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight", + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias", + "onnx::MatMul_17337_11_0_101_12_23_0", + "onnx::MatMul_17337_11_0_101_12_23_1", + "onnx::MatMul_17338_11_0_102", + "onnx::MatMul_17339_11_0_103", + "GroupNorm_16_wts_6_0_12", + "NhwcConv_17_weight_NHWC", + "GroupNorm_17_wts_6_0_13", + "NhwcConv_18_weight_NHWC", + "GroupNorm_18_wts_6_1_4", + "onnx::MatMul_17346_11_0_104", + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias", + "Attention_24_qkv_weight_key", + "Attention_24_qkv_weight_query", + "Attention_24_qkv_weight_value", + "onnx::MatMul_17357_11_0_105", + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_17358", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts", + "onnx::MatMul_17368_11_0_106", + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_17369_11_0_107_12_24_0", + "onnx::MatMul_17369_11_0_107_12_24_1", + "onnx::MatMul_17370_11_0_108", + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias", + "Attention_25_qkv_weight_key", + "Attention_25_qkv_weight_query", + "Attention_25_qkv_weight_value", + "onnx::MatMul_17381_11_0_109", + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias", + "onnx::MatMul_17382", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts", + "onnx::MatMul_17392_11_0_110", + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias", + "onnx::MatMul_17393_11_0_111_12_25_0", + "onnx::MatMul_17393_11_0_111_12_25_1", + "onnx::MatMul_17394_11_0_112", + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias", + "Attention_26_qkv_weight_key", + "Attention_26_qkv_weight_query", + "Attention_26_qkv_weight_value", + "onnx::MatMul_17405_11_0_113", + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias", + "onnx::MatMul_17406", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts", + "onnx::MatMul_17416_11_0_114", + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias", + "onnx::MatMul_17417_11_0_115_12_26_0", + "onnx::MatMul_17417_11_0_115_12_26_1", + "onnx::MatMul_17418_11_0_116", + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias", + "Attention_27_qkv_weight_key", + "Attention_27_qkv_weight_query", + "Attention_27_qkv_weight_value", + "onnx::MatMul_17429_11_0_117", + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias", + "onnx::MatMul_17430", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts", + "onnx::MatMul_17440_11_0_118", + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias", + "onnx::MatMul_17441_11_0_119_12_27_0", + "onnx::MatMul_17441_11_0_119_12_27_1", + "onnx::MatMul_17442_11_0_120", + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias", + "Attention_28_qkv_weight_key", + "Attention_28_qkv_weight_query", + "Attention_28_qkv_weight_value", + "onnx::MatMul_17453_11_0_121", + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias", + "onnx::MatMul_17454", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts", + "onnx::MatMul_17464_11_0_122", + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias", + "onnx::MatMul_17465_11_0_123_12_28_0", + "onnx::MatMul_17465_11_0_123_12_28_1", + "onnx::MatMul_17466_11_0_124", + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias", + "Attention_29_qkv_weight_key", + "Attention_29_qkv_weight_query", + "Attention_29_qkv_weight_value", + "onnx::MatMul_17477_11_0_125", + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias", + "onnx::MatMul_17478", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts", + "onnx::MatMul_17488_11_0_126", + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias", + "onnx::MatMul_17489_11_0_127_12_29_0", + "onnx::MatMul_17489_11_0_127_12_29_1", + "onnx::MatMul_17490_11_0_128", + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias", + "Attention_30_qkv_weight_key", + "Attention_30_qkv_weight_query", + "Attention_30_qkv_weight_value", + "onnx::MatMul_17501_11_0_129", + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias", + "onnx::MatMul_17502", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts", + "onnx::MatMul_17512_11_0_130", + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias", + "onnx::MatMul_17513_11_0_131_12_30_0", + "onnx::MatMul_17513_11_0_131_12_30_1", + "onnx::MatMul_17514_11_0_132", + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias", + "Attention_31_qkv_weight_key", + "Attention_31_qkv_weight_query", + "Attention_31_qkv_weight_value", + "onnx::MatMul_17525_11_0_133", + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias", + "onnx::MatMul_17526", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts", + "onnx::MatMul_17536_11_0_134", + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias", + "onnx::MatMul_17537_11_0_135_12_31_0", + "onnx::MatMul_17537_11_0_135_12_31_1", + "onnx::MatMul_17538_11_0_136", + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias", + "Attention_32_qkv_weight_key", + "Attention_32_qkv_weight_query", + "Attention_32_qkv_weight_value", + "onnx::MatMul_17549_11_0_137", + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias", + "onnx::MatMul_17550", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts", + "onnx::MatMul_17560_11_0_138", + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias", + "onnx::MatMul_17561_11_0_139_12_32_0", + "onnx::MatMul_17561_11_0_139_12_32_1", + "onnx::MatMul_17562_11_0_140", + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias", + "Attention_33_qkv_weight_key", + "Attention_33_qkv_weight_query", + "Attention_33_qkv_weight_value", + "onnx::MatMul_17573_11_0_141", + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias", + "onnx::MatMul_17574", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts", + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts", + "onnx::MatMul_17584_11_0_142", + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight", + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias", + "onnx::MatMul_17585_11_0_143_12_33_0", + "onnx::MatMul_17585_11_0_143_12_33_1", + "onnx::MatMul_17586_11_0_144", + "onnx::MatMul_17587_11_0_145", + "GroupNorm_19_wts_6_0_14", + "NhwcConv_19_weight_NHWC", + "GroupNorm_20_wts_6_0_15", + "NhwcConv_20_weight_NHWC", + "GroupNorm_21_wts_6_0_16", + "NhwcConv_21_weight_NHWC", + "GroupNorm_22_wts_6_0_17", + "NhwcConv_22_weight_NHWC", + "NhwcConv_23_weight_NHWC", + "GroupNorm_23_wts_6_1_5", + "onnx::MatMul_17598_11_0_146", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias", + "Attention_34_qkv_weight_key", + "Attention_34_qkv_weight_query", + "Attention_34_qkv_weight_value", + "onnx::MatMul_17609_11_0_147", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_17610", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts", + "onnx::MatMul_17620_11_0_148", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_17621_11_0_149_12_34_0", + "onnx::MatMul_17621_11_0_149_12_34_1", + "onnx::MatMul_17622_11_0_150", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias", + "Attention_35_qkv_weight_key", + "Attention_35_qkv_weight_query", + "Attention_35_qkv_weight_value", + "onnx::MatMul_17633_11_0_151", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias", + "onnx::MatMul_17634", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts", + "onnx::MatMul_17644_11_0_152", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias", + "onnx::MatMul_17645_11_0_153_12_35_0", + "onnx::MatMul_17645_11_0_153_12_35_1", + "onnx::MatMul_17646_11_0_154", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias", + "Attention_36_qkv_weight_key", + "Attention_36_qkv_weight_query", + "Attention_36_qkv_weight_value", + "onnx::MatMul_17657_11_0_155", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias", + "onnx::MatMul_17658", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts", + "onnx::MatMul_17668_11_0_156", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias", + "onnx::MatMul_17669_11_0_157_12_36_0", + "onnx::MatMul_17669_11_0_157_12_36_1", + "onnx::MatMul_17670_11_0_158", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias", + "Attention_37_qkv_weight_key", + "Attention_37_qkv_weight_query", + "Attention_37_qkv_weight_value", + "onnx::MatMul_17681_11_0_159", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias", + "onnx::MatMul_17682", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts", + "onnx::MatMul_17692_11_0_160", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias", + "onnx::MatMul_17693_11_0_161_12_37_0", + "onnx::MatMul_17693_11_0_161_12_37_1", + "onnx::MatMul_17694_11_0_162", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias", + "Attention_38_qkv_weight_key", + "Attention_38_qkv_weight_query", + "Attention_38_qkv_weight_value", + "onnx::MatMul_17705_11_0_163", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias", + "onnx::MatMul_17706", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts", + "onnx::MatMul_17716_11_0_164", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias", + "onnx::MatMul_17717_11_0_165_12_38_0", + "onnx::MatMul_17717_11_0_165_12_38_1", + "onnx::MatMul_17718_11_0_166", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias", + "Attention_39_qkv_weight_key", + "Attention_39_qkv_weight_query", + "Attention_39_qkv_weight_value", + "onnx::MatMul_17729_11_0_167", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias", + "onnx::MatMul_17730", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts", + "onnx::MatMul_17740_11_0_168", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias", + "onnx::MatMul_17741_11_0_169_12_39_0", + "onnx::MatMul_17741_11_0_169_12_39_1", + "onnx::MatMul_17742_11_0_170", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias", + "Attention_40_qkv_weight_key", + "Attention_40_qkv_weight_query", + "Attention_40_qkv_weight_value", + "onnx::MatMul_17753_11_0_171", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias", + "onnx::MatMul_17754", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts", + "onnx::MatMul_17764_11_0_172", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias", + "onnx::MatMul_17765_11_0_173_12_40_0", + "onnx::MatMul_17765_11_0_173_12_40_1", + "onnx::MatMul_17766_11_0_174", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias", + "Attention_41_qkv_weight_key", + "Attention_41_qkv_weight_query", + "Attention_41_qkv_weight_value", + "onnx::MatMul_17777_11_0_175", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias", + "onnx::MatMul_17778", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts", + "onnx::MatMul_17788_11_0_176", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias", + "onnx::MatMul_17789_11_0_177_12_41_0", + "onnx::MatMul_17789_11_0_177_12_41_1", + "onnx::MatMul_17790_11_0_178", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias", + "Attention_42_qkv_weight_key", + "Attention_42_qkv_weight_query", + "Attention_42_qkv_weight_value", + "onnx::MatMul_17801_11_0_179", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias", + "onnx::MatMul_17802", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts", + "onnx::MatMul_17812_11_0_180", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias", + "onnx::MatMul_17813_11_0_181_12_42_0", + "onnx::MatMul_17813_11_0_181_12_42_1", + "onnx::MatMul_17814_11_0_182", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias", + "Attention_43_qkv_weight_key", + "Attention_43_qkv_weight_query", + "Attention_43_qkv_weight_value", + "onnx::MatMul_17825_11_0_183", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias", + "onnx::MatMul_17826", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts", + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts", + "onnx::MatMul_17836_11_0_184", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias", + "onnx::MatMul_17837_11_0_185_12_43_0", + "onnx::MatMul_17837_11_0_185_12_43_1", + "onnx::MatMul_17838_11_0_186", + "onnx::MatMul_17839_11_0_187", + "GroupNorm_24_wts_6_0_18", + "NhwcConv_24_weight_NHWC", + "GroupNorm_25_wts_6_0_19", + "NhwcConv_25_weight_NHWC", + "NhwcConv_26_weight_NHWC", + "GroupNorm_26_wts_6_1_6", + "onnx::MatMul_17846_11_0_188", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias", + "Attention_44_qkv_weight_key", + "Attention_44_qkv_weight_query", + "Attention_44_qkv_weight_value", + "onnx::MatMul_17857_11_0_189", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_17858", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts", + "onnx::MatMul_17868_11_0_190", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_17869_11_0_191_12_44_0", + "onnx::MatMul_17869_11_0_191_12_44_1", + "onnx::MatMul_17870_11_0_192", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias", + "Attention_45_qkv_weight_key", + "Attention_45_qkv_weight_query", + "Attention_45_qkv_weight_value", + "onnx::MatMul_17881_11_0_193", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias", + "onnx::MatMul_17882", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts", + "onnx::MatMul_17892_11_0_194", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias", + "onnx::MatMul_17893_11_0_195_12_45_0", + "onnx::MatMul_17893_11_0_195_12_45_1", + "onnx::MatMul_17894_11_0_196", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias", + "Attention_46_qkv_weight_key", + "Attention_46_qkv_weight_query", + "Attention_46_qkv_weight_value", + "onnx::MatMul_17905_11_0_197", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias", + "onnx::MatMul_17906", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts", + "onnx::MatMul_17916_11_0_198", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias", + "onnx::MatMul_17917_11_0_199_12_46_0", + "onnx::MatMul_17917_11_0_199_12_46_1", + "onnx::MatMul_17918_11_0_200", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias", + "Attention_47_qkv_weight_key", + "Attention_47_qkv_weight_query", + "Attention_47_qkv_weight_value", + "onnx::MatMul_17929_11_0_201", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias", + "onnx::MatMul_17930", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts", + "onnx::MatMul_17940_11_0_202", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias", + "onnx::MatMul_17941_11_0_203_12_47_0", + "onnx::MatMul_17941_11_0_203_12_47_1", + "onnx::MatMul_17942_11_0_204", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias", + "Attention_48_qkv_weight_key", + "Attention_48_qkv_weight_query", + "Attention_48_qkv_weight_value", + "onnx::MatMul_17953_11_0_205", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias", + "onnx::MatMul_17954", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts", + "onnx::MatMul_17964_11_0_206", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias", + "onnx::MatMul_17965_11_0_207_12_48_0", + "onnx::MatMul_17965_11_0_207_12_48_1", + "onnx::MatMul_17966_11_0_208", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias", + "Attention_49_qkv_weight_key", + "Attention_49_qkv_weight_query", + "Attention_49_qkv_weight_value", + "onnx::MatMul_17977_11_0_209", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias", + "onnx::MatMul_17978", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts", + "onnx::MatMul_17988_11_0_210", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias", + "onnx::MatMul_17989_11_0_211_12_49_0", + "onnx::MatMul_17989_11_0_211_12_49_1", + "onnx::MatMul_17990_11_0_212", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias", + "Attention_50_qkv_weight_key", + "Attention_50_qkv_weight_query", + "Attention_50_qkv_weight_value", + "onnx::MatMul_18001_11_0_213", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias", + "onnx::MatMul_18002", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts", + "onnx::MatMul_18012_11_0_214", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias", + "onnx::MatMul_18013_11_0_215_12_50_0", + "onnx::MatMul_18013_11_0_215_12_50_1", + "onnx::MatMul_18014_11_0_216", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias", + "Attention_51_qkv_weight_key", + "Attention_51_qkv_weight_query", + "Attention_51_qkv_weight_value", + "onnx::MatMul_18025_11_0_217", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias", + "onnx::MatMul_18026", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts", + "onnx::MatMul_18036_11_0_218", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias", + "onnx::MatMul_18037_11_0_219_12_51_0", + "onnx::MatMul_18037_11_0_219_12_51_1", + "onnx::MatMul_18038_11_0_220", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias", + "Attention_52_qkv_weight_key", + "Attention_52_qkv_weight_query", + "Attention_52_qkv_weight_value", + "onnx::MatMul_18049_11_0_221", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias", + "onnx::MatMul_18050", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts", + "onnx::MatMul_18060_11_0_222", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias", + "onnx::MatMul_18061_11_0_223_12_52_0", + "onnx::MatMul_18061_11_0_223_12_52_1", + "onnx::MatMul_18062_11_0_224", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias", + "Attention_53_qkv_weight_key", + "Attention_53_qkv_weight_query", + "Attention_53_qkv_weight_value", + "onnx::MatMul_18073_11_0_225", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias", + "onnx::MatMul_18074", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts", + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts", + "onnx::MatMul_18084_11_0_226", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias", + "onnx::MatMul_18085_11_0_227_12_53_0", + "onnx::MatMul_18085_11_0_227_12_53_1", + "onnx::MatMul_18086_11_0_228", + "onnx::MatMul_18087_11_0_229", + "GroupNorm_27_wts_6_0_20", + "NhwcConv_27_weight_NHWC", + "GroupNorm_28_wts_6_0_21", + "NhwcConv_28_weight_NHWC", + "NhwcConv_29_weight_NHWC", + "GroupNorm_29_wts_6_1_7", + "onnx::MatMul_18094_11_0_230", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias", + "Attention_54_qkv_weight_key", + "Attention_54_qkv_weight_query", + "Attention_54_qkv_weight_value", + "onnx::MatMul_18105_11_0_231", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias", + "onnx::MatMul_18106", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts", + "onnx::MatMul_18116_11_0_232", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias", + "onnx::MatMul_18117_11_0_233_12_54_0", + "onnx::MatMul_18117_11_0_233_12_54_1", + "onnx::MatMul_18118_11_0_234", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias", + "Attention_55_qkv_weight_key", + "Attention_55_qkv_weight_query", + "Attention_55_qkv_weight_value", + "onnx::MatMul_18129_11_0_235", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias", + "onnx::MatMul_18130", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts", + "onnx::MatMul_18140_11_0_236", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias", + "onnx::MatMul_18141_11_0_237_12_55_0", + "onnx::MatMul_18141_11_0_237_12_55_1", + "onnx::MatMul_18142_11_0_238", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias", + "Attention_56_qkv_weight_key", + "Attention_56_qkv_weight_query", + "Attention_56_qkv_weight_value", + "onnx::MatMul_18153_11_0_239", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias", + "onnx::MatMul_18154", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts", + "onnx::MatMul_18164_11_0_240", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias", + "onnx::MatMul_18165_11_0_241_12_56_0", + "onnx::MatMul_18165_11_0_241_12_56_1", + "onnx::MatMul_18166_11_0_242", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias", + "Attention_57_qkv_weight_key", + "Attention_57_qkv_weight_query", + "Attention_57_qkv_weight_value", + "onnx::MatMul_18177_11_0_243", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias", + "onnx::MatMul_18178", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts", + "onnx::MatMul_18188_11_0_244", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias", + "onnx::MatMul_18189_11_0_245_12_57_0", + "onnx::MatMul_18189_11_0_245_12_57_1", + "onnx::MatMul_18190_11_0_246", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias", + "Attention_58_qkv_weight_key", + "Attention_58_qkv_weight_query", + "Attention_58_qkv_weight_value", + "onnx::MatMul_18201_11_0_247", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias", + "onnx::MatMul_18202", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts", + "onnx::MatMul_18212_11_0_248", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias", + "onnx::MatMul_18213_11_0_249_12_58_0", + "onnx::MatMul_18213_11_0_249_12_58_1", + "onnx::MatMul_18214_11_0_250", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias", + "Attention_59_qkv_weight_key", + "Attention_59_qkv_weight_query", + "Attention_59_qkv_weight_value", + "onnx::MatMul_18225_11_0_251", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias", + "onnx::MatMul_18226", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts", + "onnx::MatMul_18236_11_0_252", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias", + "onnx::MatMul_18237_11_0_253_12_59_0", + "onnx::MatMul_18237_11_0_253_12_59_1", + "onnx::MatMul_18238_11_0_254", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias", + "Attention_60_qkv_weight_key", + "Attention_60_qkv_weight_query", + "Attention_60_qkv_weight_value", + "onnx::MatMul_18249_11_0_255", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias", + "onnx::MatMul_18250", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts", + "onnx::MatMul_18260_11_0_256", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias", + "onnx::MatMul_18261_11_0_257_12_60_0", + "onnx::MatMul_18261_11_0_257_12_60_1", + "onnx::MatMul_18262_11_0_258", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias", + "Attention_61_qkv_weight_key", + "Attention_61_qkv_weight_query", + "Attention_61_qkv_weight_value", + "onnx::MatMul_18273_11_0_259", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias", + "onnx::MatMul_18274", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts", + "onnx::MatMul_18284_11_0_260", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias", + "onnx::MatMul_18285_11_0_261_12_61_0", + "onnx::MatMul_18285_11_0_261_12_61_1", + "onnx::MatMul_18286_11_0_262", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias", + "Attention_62_qkv_weight_key", + "Attention_62_qkv_weight_query", + "Attention_62_qkv_weight_value", + "onnx::MatMul_18297_11_0_263", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias", + "onnx::MatMul_18298", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts", + "onnx::MatMul_18308_11_0_264", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias", + "onnx::MatMul_18309_11_0_265_12_62_0", + "onnx::MatMul_18309_11_0_265_12_62_1", + "onnx::MatMul_18310_11_0_266", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias", + "Attention_63_qkv_weight_key", + "Attention_63_qkv_weight_query", + "Attention_63_qkv_weight_value", + "onnx::MatMul_18321_11_0_267", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias", + "onnx::MatMul_18322", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts", + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts", + "onnx::MatMul_18332_11_0_268", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight", + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias", + "onnx::MatMul_18333_11_0_269_12_63_0", + "onnx::MatMul_18333_11_0_269_12_63_1", + "onnx::MatMul_18334_11_0_270", + "onnx::MatMul_18335_11_0_271", + "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0", + "NhwcConv_30_weight_NHWC", + "GroupNorm_30_wts_6_0_22", + "NhwcConv_31_weight_NHWC", + "GroupNorm_31_wts_6_0_23", + "NhwcConv_32_weight_NHWC", + "NhwcConv_33_weight_NHWC", + "GroupNorm_32_wts_6_1_8", + "onnx::MatMul_18343_11_0_272", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", + "Attention_64_qkv_weight_key", + "Attention_64_qkv_weight_query", + "Attention_64_qkv_weight_value", + "onnx::MatMul_18354_11_0_273", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_18355", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts", + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts", + "onnx::MatMul_18365_11_0_274", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_18366_11_0_275_12_64_0", + "onnx::MatMul_18366_11_0_275_12_64_1", + "onnx::MatMul_18367_11_0_276", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias", + "Attention_65_qkv_weight_key", + "Attention_65_qkv_weight_query", + "Attention_65_qkv_weight_value", + "onnx::MatMul_18378_11_0_277", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias", + "onnx::MatMul_18379", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts", + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts", + "onnx::MatMul_18389_11_0_278", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias", + "onnx::MatMul_18390_11_0_279_12_65_0", + "onnx::MatMul_18390_11_0_279_12_65_1", + "onnx::MatMul_18391_11_0_280", + "onnx::MatMul_18392_11_0_281", + "GroupNorm_33_wts_6_0_24", + "NhwcConv_34_weight_NHWC", + "GroupNorm_34_wts_6_0_25", + "NhwcConv_35_weight_NHWC", + "NhwcConv_36_weight_NHWC", + "GroupNorm_35_wts_6_1_9", + "onnx::MatMul_18399_11_0_282", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", + "Attention_66_qkv_weight_key", + "Attention_66_qkv_weight_query", + "Attention_66_qkv_weight_value", + "onnx::MatMul_18410_11_0_283", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_18411", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts", + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts", + "onnx::MatMul_18421_11_0_284", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_18422_11_0_285_12_66_0", + "onnx::MatMul_18422_11_0_285_12_66_1", + "onnx::MatMul_18423_11_0_286", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias", + "Attention_67_qkv_weight_key", + "Attention_67_qkv_weight_query", + "Attention_67_qkv_weight_value", + "onnx::MatMul_18434_11_0_287", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias", + "onnx::MatMul_18435", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts", + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts", + "onnx::MatMul_18445_11_0_288", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias", + "onnx::MatMul_18446_11_0_289_12_67_0", + "onnx::MatMul_18446_11_0_289_12_67_1", + "onnx::MatMul_18447_11_0_290", + "onnx::MatMul_18448_11_0_291", + "GroupNorm_36_wts_6_0_26", + "NhwcConv_37_weight_NHWC", + "GroupNorm_37_wts_6_0_27", + "NhwcConv_38_weight_NHWC", + "NhwcConv_39_weight_NHWC", + "GroupNorm_38_wts_6_1_10", + "onnx::MatMul_18455_11_0_292", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias", + "Attention_68_qkv_weight_key", + "Attention_68_qkv_weight_query", + "Attention_68_qkv_weight_value", + "onnx::MatMul_18466_11_0_293", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias", + "onnx::MatMul_18467", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts", + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts", + "onnx::MatMul_18477_11_0_294", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias", + "onnx::MatMul_18478_11_0_295_12_68_0", + "onnx::MatMul_18478_11_0_295_12_68_1", + "onnx::MatMul_18479_11_0_296", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias", + "Attention_69_qkv_weight_key", + "Attention_69_qkv_weight_query", + "Attention_69_qkv_weight_value", + "onnx::MatMul_18490_11_0_297", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias", + "onnx::MatMul_18491", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts", + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts", + "onnx::MatMul_18501_11_0_298", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight", + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias", + "onnx::MatMul_18502_11_0_299_12_69_0", + "onnx::MatMul_18502_11_0_299_12_69_1", + "onnx::MatMul_18503_11_0_300", + "onnx::MatMul_18504_11_0_301", + "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1", + "NhwcConv_40_weight_NHWC", + "GroupNorm_39_wts_6_0_28", + "NhwcConv_41_weight_NHWC", + "GroupNorm_40_wts_6_0_29", + "NhwcConv_42_weight_NHWC", + "NhwcConv_43_weight_NHWC", + "GroupNorm_41_wts_6_0_30", + "NhwcConv_44_weight_NHWC", + "GroupNorm_42_wts_6_0_31", + "NhwcConv_45_weight_NHWC", + "NhwcConv_46_weight_NHWC", + "GroupNorm_43_wts_6_0_32", + "NhwcConv_47_weight_NHWC", + "GroupNorm_44_wts_6_0_33", + "NhwcConv_48_weight_NHWC", + "NhwcConv_49_weight_NHWC", + "GroupNorm_45_wts_6_0_34", + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts", + "NhwcConv_50_weight_NHWC" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "Transpose_34_out-sample.out0_1_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 128, + 128, + 4 + ], + "size_in_bytes": 262144, + "op_tensor_size": 262144, + "offset": 0 + }, + "/unet/Concat_3_output_0.out2_3_1": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 2816 + ], + "size_in_bytes": 11264, + "op_tensor_size": 11264, + "offset": 262144 + }, + "encoder_hidden_states.out2_6_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 2048 + ], + "size_in_bytes": 630784, + "op_tensor_size": 630784, + "offset": 87953664 + }, + "/unet/time_proj/Concat_1_output_0.out2_3_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 43166720 + }, + "NhwcConv_50_out-out_sample.out0_1_50": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 2, + 128, + 128, + 4 + ], + "size_in_bytes": 262144, + "op_tensor_size": 262144, + "offset": 0 + }, + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out0_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 0 + }, + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.out7_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 20971520 + }, + "GroupNorm_0.out7_0.out6_0_0_bfp.out7_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 31457280 + }, + "/unet/add_embedding/linear_1/Gemm_output_0.out2_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 41943040 + }, + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 41948160 + }, + "/unet/add_embedding/linear_2/Gemm_output_0.out2_3_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 41953280 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 41958400 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 42155520 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 42352640 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 42549760 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 42746880 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 42944000 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 43141120 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 43338240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 43535360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 43929600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 44323840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 44718080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 45112320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 45506560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 45900800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 46295040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 46689280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 47083520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 47477760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 47872000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 48266240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 48660480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 49054720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 49448960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 49843200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 50237440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 50631680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 51025920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 51420160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 51814400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 52208640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 52602880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 52997120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 53391360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 53785600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 54179840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 54574080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 54968320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 55362560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 55756800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 56151040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 56545280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 56939520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 57333760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 57728000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 58122240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 58516480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 58910720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 59304960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 59699200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 60093440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 60487680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 60881920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 61276160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 61670400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 62064640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 62458880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 62853120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 63247360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 63641600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 64035840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 64430080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 64824320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 65218560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 65612800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 66007040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 66401280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 66795520 + }, + "/unet/time_embedding/linear_1/Gemm_output_0.out2_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 67189760 + }, + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid_out.out14_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 67194880 + }, + "/unet/time_embedding/linear_2/Gemm_output_0.out2_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 67200000 + }, + "/unet/Add_output_0.out_17_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 67205120 + }, + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid_out.out14_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 67210240 + }, + "/unet/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 67215360 + }, + "NhwcConv_1_out-/unet/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1_bfp.out1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 67216640 + }, + "/unet/down_blocks.0/resnets.0/Add.out_17_1_1_bfp.out18_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 77702400 + }, + "GroupNorm_1.out7_1.out6_0_1_bfp.out7_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 88188160 + }, + "NhwcConv_2_out-/unet/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2_bfp.out1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 98673920 + }, + "/unet/down_blocks.0/resnets.0/Add_1.out_17_1_2_bfp.out18_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 109159680 + }, + "GroupNorm_2.out7_2.out6_0_2_bfp.out7_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 119645440 + }, + "/unet/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 130131200 + }, + "NhwcConv_3_out-/unet/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_3_bfp.out1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 130132480 + }, + "/unet/down_blocks.0/resnets.1/Add.out_17_1_3_bfp.out18_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 140618240 + }, + "GroupNorm_3.out7_3.out6_0_3_bfp.out7_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 151104000 + }, + "NhwcConv_4_out-/unet/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_4_bfp.out1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 161589760 + }, + "/unet/down_blocks.0/resnets.1/Add_1.out_17_1_4_bfp.out18_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 172075520 + }, + "NhwcConv_5_out-/unet/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_5_bfp.out1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 182561280 + }, + "GroupNorm_4.out7_4.out6_0_4_bfp.out7_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 185182720 + }, + "/unet/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 187804160 + }, + "NhwcConv_6_out-/unet/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_7_bfp.out1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 187806720 + }, + "/unet/down_blocks.1/resnets.0/Add.out_17_1_5_bfp.out18_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 193049600 + }, + "GroupNorm_5.out7_5.out6_0_5_bfp.out7_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 198292480 + }, + "/unet/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 203535360 + }, + "/unet/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 203537920 + }, + "/unet/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 203543040 + }, + "/unet/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 203548160 + }, + "/unet/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 203553280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 203558400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 203952640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 204346880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 204741120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 205135360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 205529600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 205923840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 206318080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 206712320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 207106560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 207500800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 207895040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 208289280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 208683520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 209077760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 209472000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 209866240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 210260480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 210654720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 211048960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 211443200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 211837440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 212231680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 212625920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 213020160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 213414400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 213808640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 214202880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 214597120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 214991360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_98": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 215385600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_99": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 215779840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_100": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 216174080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_101": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 216568320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_102": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 216962560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_103": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 217356800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_104": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 217751040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 218145280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_106": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 218539520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_107": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 218933760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_108": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 219328000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_109": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 219722240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_110": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 220116480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 220510720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_k/MatMul_output_0.out2_6_112": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 220904960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_v/MatMul_output_0.out2_6_113": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 221299200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_k/MatMul_output_0.out2_6_114": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 221693440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_v/MatMul_output_0.out2_6_115": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 222087680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_k/MatMul_output_0.out2_6_116": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 222481920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_v/MatMul_output_0.out2_6_117": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 222876160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_k/MatMul_output_0.out2_6_118": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 223270400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_v/MatMul_output_0.out2_6_119": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 223664640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_k/MatMul_output_0.out2_6_120": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 224058880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_v/MatMul_output_0.out2_6_121": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 224453120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_k/MatMul_output_0.out2_6_122": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 224847360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_v/MatMul_output_0.out2_6_123": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 225241600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_k/MatMul_output_0.out2_6_124": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 225635840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_v/MatMul_output_0.out2_6_125": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 226030080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_k/MatMul_output_0.out2_6_126": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 226424320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_v/MatMul_output_0.out2_6_127": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 77, + 64 + ], + "size_in_bytes": 394240, + "op_tensor_size": 394240, + "offset": 226818560 + }, + "/unet/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 227212800 + }, + "/unet/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 227217920 + }, + "/unet/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 227223040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_128": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 227228160 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_129": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 227425280 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_130": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 227622400 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_131": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 227819520 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_132": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 228016640 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_133": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 228213760 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_134": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 228410880 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_135": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 228608000 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_136": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 228805120 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_137": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 229002240 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_k/MatMul_output_0.out2_6_138": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 229199360 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_v/MatMul_output_0.out2_6_139": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 77, + 64 + ], + "size_in_bytes": 197120, + "op_tensor_size": 197120, + "offset": 229396480 + }, + "/unet/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 229593600 + }, + "/unet/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 229596160 + }, + "/unet/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 229598720 + }, + "/unet/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 229601280 + }, + "/unet/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 229602560 + }, + "/unet/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 229603840 + }, + "NhwcConv_7_out-/unet/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_8_bfp.out1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 229605120 + }, + "NhwcConv_8_out-/unet/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_6_bfp.out1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 234848000 + }, + "/unet/down_blocks.1/resnets.0/Add_1.out_17_1_6_bfp.out18_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 240090880 + }, + "/unet/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_0_bfp.out7_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 245333760 + }, + "/unet/down_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_4_bfp.out9_231": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 250576640 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_0_bfp.out13_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 255819520 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_0_bfp.out9_233": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 261062400 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_1_bfp.out9_232": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 266305280 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_2_bfp.out9_234": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 271548160 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0_bfp.out11_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 276791040 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_5_bfp.out9_235": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 282033920 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_7_bfp.out18_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 287276800 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_1_bfp.out13_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 292519680 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 297762560 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 303005440 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 313491200 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.out9_236": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 323976960 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_6_bfp.out9_236": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 329219840 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_8_bfp.out18_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 334462720 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_2_bfp.out13_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 339705600 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0_bfp.out9_237": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 344948480 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_7_bfp.out9_238": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 365920000 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_2_bfp.out22_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 386891520 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_8_bfp.out9_239": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 407863040 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_9_bfp.out18_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 413105920 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_3_bfp.out13_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 418348800 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_3_bfp.out9_241": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 423591680 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_4_bfp.out9_240": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 428834560 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_5_bfp.out9_242": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 434077440 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_2_bfp.out11_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 439320320 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_9_bfp.out9_243": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 444563200 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_10_bfp.out18_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 449806080 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_4_bfp.out13_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 455048960 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 460291840 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 465534720 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 476020480 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.out9_244": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 486506240 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_10_bfp.out9_244": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 491749120 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_11_bfp.out18_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 496992000 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_5_bfp.out13_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 502234880 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_1_bfp.out9_245": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 507477760 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_11_bfp.out9_246": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 528449280 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_3_bfp.out22_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 549420800 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_12_bfp.out9_247": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 570392320 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_12_bfp.out18_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 575635200 + }, + "/unet/down_blocks.1/attentions.0/Reshape_1_output_0.out2_0_17_bfp.out9_248": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 580878080 + }, + "/unet/down_blocks.1/attentions.0/Add.out_17_1_13_bfp.out18_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 586120960 + }, + "GroupNorm_7.out7_7.out6_0_6_bfp.out7_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 591363840 + }, + "NhwcConv_9_out-/unet/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_9_bfp.out1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 596606720 + }, + "/unet/down_blocks.1/resnets.1/Add.out_17_1_14_bfp.out18_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 601849600 + }, + "GroupNorm_8.out7_8.out6_0_7_bfp.out7_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 607092480 + }, + "NhwcConv_10_out-/unet/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_10_bfp.out1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 612335360 + }, + "/unet/down_blocks.1/resnets.1/Add_1.out_17_1_15_bfp.out18_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 617578240 + }, + "/unet/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_1_bfp.out7_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 622821120 + }, + "/unet/down_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_13_bfp.out9_249": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 628064000 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_6_bfp.out13_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 633306880 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_6_bfp.out9_251": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 638549760 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_7_bfp.out9_250": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 643792640 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_8_bfp.out9_252": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 649035520 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4_bfp.out11_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 654278400 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14_bfp.out9_253": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 659521280 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_16_bfp.out18_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 664764160 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_7_bfp.out13_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 670007040 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 675249920 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 680492800 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 690978560 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.out9_254": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 701464320 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15_bfp.out9_254": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 706707200 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_17_bfp.out18_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 711950080 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_8_bfp.out13_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 717192960 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2_bfp.out9_255": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 722435840 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16_bfp.out9_256": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 743407360 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_4_bfp.out22_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 764378880 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17_bfp.out9_257": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 785350400 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_18_bfp.out18_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 790593280 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_9_bfp.out13_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 795836160 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_9_bfp.out9_259": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 801079040 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_10_bfp.out9_258": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 806321920 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_11_bfp.out9_260": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 811564800 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_6_bfp.out11_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 816807680 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_18_bfp.out9_261": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 822050560 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_19_bfp.out18_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 827293440 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_10_bfp.out13_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 832536320 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 837779200 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 843022080 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 853507840 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.out9_262": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 863993600 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_19_bfp.out9_262": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 869236480 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_20_bfp.out18_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 874479360 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_11_bfp.out13_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 879722240 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_3_bfp.out9_263": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 884965120 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_20_bfp.out9_264": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 905936640 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_5_bfp.out22_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 926908160 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_21_bfp.out9_265": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 947879680 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_21_bfp.out18_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 953122560 + }, + "/unet/down_blocks.1/attentions.1/Reshape_1_output_0.out2_0_18_bfp.out9_266": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 958365440 + }, + "/unet/down_blocks.1/attentions.1/Add.out_17_1_22_bfp.out18_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 963608320 + }, + "NhwcConv_11_out-/unet/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_11_bfp.out1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 1310720, + "op_tensor_size": 1310720, + "offset": 968851200 + }, + "GroupNorm_10.out7_10.out6_0_8_bfp.out7_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 1310720, + "op_tensor_size": 1310720, + "offset": 970161920 + }, + "NhwcConv_12_out-/unet/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_13_bfp.out1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 971472640 + }, + "/unet/down_blocks.2/resnets.0/Add.out_17_1_23_bfp.out18_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 974094080 + }, + "GroupNorm_11.out7_11.out6_0_9_bfp.out7_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 976715520 + }, + "NhwcConv_13_out-/unet/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_14_bfp.out1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 979336960 + }, + "NhwcConv_14_out-/unet/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_12_bfp.out1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 981958400 + }, + "/unet/down_blocks.2/resnets.0/Add_1.out_17_1_24_bfp.out18_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 984579840 + }, + "/unet/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out6_1_2_bfp.out7_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 987201280 + }, + "/unet/down_blocks.2/attentions.0/proj_in/Add_output_0.out2_3_22_bfp.out9_267": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 989822720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_12_bfp.out13_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 992444160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_12_bfp.out9_269": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 995065600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_13_bfp.out9_268": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 997687040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_14_bfp.out9_270": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1000308480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8_bfp.out11_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1002929920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_23_bfp.out9_271": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1005551360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_25_bfp.out18_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1008172800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_13_bfp.out13_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1010794240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1013415680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1016037120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1021280000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.out9_272": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1026522880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_24_bfp.out9_272": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1029144320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_26_bfp.out18_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1031765760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_14_bfp.out13_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1034387200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4_bfp.out9_273": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1037008640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_25_bfp.out9_274": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1047494400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_6_bfp.out22_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1057980160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_26_bfp.out9_275": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1068465920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_27_bfp.out18_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1071087360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_15_bfp.out13_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1073708800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_15_bfp.out9_277": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1076330240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_16_bfp.out9_276": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1078951680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_17_bfp.out9_278": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1081573120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_10_bfp.out11_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1084194560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_27_bfp.out9_279": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1086816000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_28_bfp.out18_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1089437440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_16_bfp.out13_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1092058880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1094680320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1097301760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1102544640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.out9_280": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1107787520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_28_bfp.out9_280": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1110408960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_29_bfp.out18_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1113030400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_17_bfp.out13_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1115651840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_5_bfp.out9_281": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1118273280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_29_bfp.out9_282": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1128759040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_7_bfp.out22_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1139244800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_30_bfp.out9_283": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1149730560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_30_bfp.out18_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1152352000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_18_bfp.out13_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1154973440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_18_bfp.out9_285": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1157594880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_19_bfp.out9_284": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1160216320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_20_bfp.out9_286": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1162837760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_12_bfp.out11_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1165459200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_31_bfp.out9_287": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1168080640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_31_bfp.out18_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1170702080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_19_bfp.out13_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1173323520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1175944960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1178566400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1183809280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.out9_288": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1189052160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_32_bfp.out9_288": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1191673600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_32_bfp.out18_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1194295040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_20_bfp.out13_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1196916480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_6_bfp.out9_289": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1199537920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_33_bfp.out9_290": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1210023680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_8_bfp.out22_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1220509440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_34_bfp.out9_291": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1230995200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_33_bfp.out18_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1233616640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_21_bfp.out13_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1236238080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_21_bfp.out9_293": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1238859520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_22_bfp.out9_292": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1241480960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_23_bfp.out9_294": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1244102400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_14_bfp.out11_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1246723840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_35_bfp.out9_295": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1249345280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_34_bfp.out18_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1251966720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_22_bfp.out13_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1254588160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1257209600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1259831040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1265073920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.out9_296": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1270316800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_36_bfp.out9_296": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1272938240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_35_bfp.out18_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1275559680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_23_bfp.out13_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1278181120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_7_bfp.out9_297": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1280802560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_37_bfp.out9_298": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1291288320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_9_bfp.out22_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1301774080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_38_bfp.out9_299": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1312259840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_36_bfp.out18_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1314881280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_24_bfp.out13_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1317502720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_24_bfp.out9_301": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1320124160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_25_bfp.out9_300": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1322745600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_26_bfp.out9_302": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1325367040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_16_bfp.out11_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1327988480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_39_bfp.out9_303": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1330609920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_37_bfp.out18_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1333231360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_25_bfp.out13_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1335852800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1338474240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1341095680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1346338560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.out9_304": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1351581440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_40_bfp.out9_304": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1354202880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_38_bfp.out18_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1356824320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_26_bfp.out13_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1359445760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_8_bfp.out9_305": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1362067200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_41_bfp.out9_306": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1372552960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_10_bfp.out22_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1383038720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_42_bfp.out9_307": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1393524480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_39_bfp.out18_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1396145920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_27_bfp.out13_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1398767360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_27_bfp.out9_309": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1401388800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_28_bfp.out9_308": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1404010240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_29_bfp.out9_310": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1406631680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_18_bfp.out11_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1409253120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_43_bfp.out9_311": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1411874560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_40_bfp.out18_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1414496000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_28_bfp.out13_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1417117440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1419738880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1422360320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1427603200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.out9_312": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1432846080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_44_bfp.out9_312": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1435467520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_41_bfp.out18_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1438088960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_29_bfp.out13_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1440710400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_9_bfp.out9_313": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1443331840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_45_bfp.out9_314": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1453817600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_11_bfp.out22_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1464303360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_46_bfp.out9_315": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1474789120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_42_bfp.out18_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1477410560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_30_bfp.out13_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1480032000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_30_bfp.out9_317": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1482653440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_31_bfp.out9_316": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1485274880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_32_bfp.out9_318": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1487896320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_20_bfp.out11_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1490517760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_47_bfp.out9_319": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1493139200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_43_bfp.out18_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1495760640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_31_bfp.out13_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1498382080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1501003520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1503624960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1508867840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.out9_320": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1514110720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_48_bfp.out9_320": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1516732160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_44_bfp.out18_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1519353600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_32_bfp.out13_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1521975040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_10_bfp.out9_321": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1524596480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_49_bfp.out9_322": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1535082240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_12_bfp.out22_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1545568000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_50_bfp.out9_323": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1556053760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_45_bfp.out18_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1558675200 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_33_bfp.out13_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1561296640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_33_bfp.out9_325": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1563918080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_34_bfp.out9_324": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1566539520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_35_bfp.out9_326": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1569160960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_22_bfp.out11_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1571782400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_51_bfp.out9_327": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1574403840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_46_bfp.out18_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1577025280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_34_bfp.out13_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1579646720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1582268160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1584889600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1590132480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.out9_328": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1595375360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_52_bfp.out9_328": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1597996800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_47_bfp.out18_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1600618240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_35_bfp.out13_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1603239680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_11_bfp.out9_329": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1605861120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_53_bfp.out9_330": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1616346880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_13_bfp.out22_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1626832640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_54_bfp.out9_331": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1637318400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_48_bfp.out18_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1639939840 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_36_bfp.out13_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1642561280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_36_bfp.out9_333": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1645182720 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_37_bfp.out9_332": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1647804160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_38_bfp.out9_334": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1650425600 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_24_bfp.out11_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1653047040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_55_bfp.out9_335": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1655668480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_49_bfp.out18_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1658289920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_37_bfp.out13_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1660911360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1663532800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1666154240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1671397120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.out9_336": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1676640000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_56_bfp.out9_336": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1679261440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_50_bfp.out18_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1681882880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_38_bfp.out13_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1684504320 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_12_bfp.out9_337": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1687125760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_57_bfp.out9_338": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1697611520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_14_bfp.out22_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1708097280 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_58_bfp.out9_339": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1718583040 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_51_bfp.out18_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1721204480 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_39_bfp.out13_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1723825920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_39_bfp.out9_341": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1726447360 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_40_bfp.out9_340": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1729068800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_41_bfp.out9_342": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1731690240 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_26_bfp.out11_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1734311680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_59_bfp.out9_343": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1736933120 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_52_bfp.out18_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1739554560 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_40_bfp.out13_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1742176000 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1744797440 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1747418880 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1752661760 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.out9_344": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1757904640 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_60_bfp.out9_344": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1760526080 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_53_bfp.out18_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1763147520 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_41_bfp.out13_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1765768960 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_13_bfp.out9_345": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1768390400 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_61_bfp.out9_346": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1778876160 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_15_bfp.out22_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1789361920 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_62_bfp.out9_347": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1799847680 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_54_bfp.out18_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1802469120 + }, + "/unet/down_blocks.2/attentions.0/Reshape_1_output_0.out2_0_19_bfp.out9_348": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1805090560 + }, + "/unet/down_blocks.2/attentions.0/Add.out_17_1_55_bfp.out18_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1807712000 + }, + "GroupNorm_13.out7_13.out6_0_10_bfp.out7_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1810333440 + }, + "NhwcConv_15_out-/unet/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_15_bfp.out1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1812954880 + }, + "/unet/down_blocks.2/resnets.1/Add.out_17_1_56_bfp.out18_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1815576320 + }, + "GroupNorm_14.out7_14.out6_0_11_bfp.out7_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1818197760 + }, + "NhwcConv_16_out-/unet/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_16_bfp.out1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1820819200 + }, + "/unet/down_blocks.2/resnets.1/Add_1.out_17_1_57_bfp.out18_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1823440640 + }, + "/unet/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out6_1_3_bfp.out7_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1826062080 + }, + "/unet/down_blocks.2/attentions.1/proj_in/Add_output_0.out2_3_63_bfp.out9_349": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1828683520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_42_bfp.out13_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1831304960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_42_bfp.out9_351": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1833926400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_43_bfp.out9_350": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1836547840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_44_bfp.out9_352": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1839169280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28_bfp.out11_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1841790720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_64_bfp.out9_353": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1844412160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_58_bfp.out18_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1847033600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_43_bfp.out13_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1849655040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1852276480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1854897920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1860140800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.out9_354": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1865383680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_65_bfp.out9_354": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1868005120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_59_bfp.out18_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1870626560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_44_bfp.out13_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1873248000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14_bfp.out9_355": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1875869440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_66_bfp.out9_356": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1886355200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_16_bfp.out22_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1896840960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_67_bfp.out9_357": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1907326720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_60_bfp.out18_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1909948160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_45_bfp.out13_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1912569600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_45_bfp.out9_359": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1915191040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_46_bfp.out9_358": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1917812480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_47_bfp.out9_360": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1920433920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_30_bfp.out11_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1923055360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_68_bfp.out9_361": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1925676800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_61_bfp.out18_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1928298240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_46_bfp.out13_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1930919680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1933541120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1936162560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 1941405440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.out9_362": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1946648320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_69_bfp.out9_362": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1949269760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_62_bfp.out18_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1951891200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_47_bfp.out13_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1954512640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_15_bfp.out9_363": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1957134080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_70_bfp.out9_364": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1967619840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_17_bfp.out22_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 1978105600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_71_bfp.out9_365": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1988591360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_63_bfp.out18_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1991212800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_48_bfp.out13_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1993834240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_48_bfp.out9_367": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1996455680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_49_bfp.out9_366": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 1999077120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_50_bfp.out9_368": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2001698560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_32_bfp.out11_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2004320000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_72_bfp.out9_369": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2006941440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_64_bfp.out18_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2009562880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_49_bfp.out13_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2012184320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2014805760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2017427200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2022670080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.out9_370": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2027912960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_73_bfp.out9_370": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2030534400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_65_bfp.out18_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2033155840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_50_bfp.out13_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2035777280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_16_bfp.out9_371": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2038398720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_74_bfp.out9_372": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2048884480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_18_bfp.out22_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2059370240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_75_bfp.out9_373": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2069856000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_66_bfp.out18_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2072477440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_51_bfp.out13_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2075098880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_51_bfp.out9_375": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2077720320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_52_bfp.out9_374": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2080341760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_53_bfp.out9_376": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2082963200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_34_bfp.out11_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2085584640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_76_bfp.out9_377": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2088206080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_67_bfp.out18_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2090827520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_52_bfp.out13_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2093448960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2096070400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2098691840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2103934720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.out9_378": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2109177600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_77_bfp.out9_378": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2111799040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_68_bfp.out18_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2114420480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_53_bfp.out13_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2117041920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_17_bfp.out9_379": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2119663360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_78_bfp.out9_380": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2130149120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_19_bfp.out22_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2140634880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_79_bfp.out9_381": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2151120640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_69_bfp.out18_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2153742080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_54_bfp.out13_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2156363520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_54_bfp.out9_383": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2158984960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_55_bfp.out9_382": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2161606400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_56_bfp.out9_384": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2164227840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_36_bfp.out11_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2166849280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_80_bfp.out9_385": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2169470720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_70_bfp.out18_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2172092160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_55_bfp.out13_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2174713600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2177335040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2179956480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2185199360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.out9_386": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2190442240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_81_bfp.out9_386": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2193063680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_71_bfp.out18_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2195685120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_56_bfp.out13_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2198306560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_18_bfp.out9_387": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2200928000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_82_bfp.out9_388": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2211413760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_20_bfp.out22_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2221899520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_83_bfp.out9_389": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2232385280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_72_bfp.out18_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2235006720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_57_bfp.out13_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2237628160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_57_bfp.out9_391": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2240249600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_58_bfp.out9_390": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2242871040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_59_bfp.out9_392": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2245492480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_38_bfp.out11_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2248113920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_84_bfp.out9_393": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2250735360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_73_bfp.out18_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2253356800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_58_bfp.out13_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2255978240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2258599680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2261221120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2266464000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.out9_394": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2271706880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_85_bfp.out9_394": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2274328320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_74_bfp.out18_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2276949760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_59_bfp.out13_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2279571200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_19_bfp.out9_395": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2282192640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_86_bfp.out9_396": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2292678400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_21_bfp.out22_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2303164160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_87_bfp.out9_397": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2313649920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_75_bfp.out18_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2316271360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_60_bfp.out13_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2318892800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_60_bfp.out9_399": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2321514240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_61_bfp.out9_398": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2324135680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_62_bfp.out9_400": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2326757120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_40_bfp.out11_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2329378560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_88_bfp.out9_401": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2332000000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_76_bfp.out18_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2334621440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_61_bfp.out13_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2337242880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2339864320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2342485760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2347728640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.out9_402": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2352971520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_89_bfp.out9_402": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2355592960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_77_bfp.out18_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2358214400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_62_bfp.out13_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2360835840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_20_bfp.out9_403": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2363457280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_90_bfp.out9_404": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2373943040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_22_bfp.out22_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2384428800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_91_bfp.out9_405": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2394914560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_78_bfp.out18_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2397536000 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_63_bfp.out13_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2400157440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_63_bfp.out9_407": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2402778880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_64_bfp.out9_406": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2405400320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_65_bfp.out9_408": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2408021760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_42_bfp.out11_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2410643200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_92_bfp.out9_409": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2413264640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_79_bfp.out18_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2415886080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_64_bfp.out13_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2418507520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2421128960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2423750400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2428993280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.out9_410": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2434236160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_93_bfp.out9_410": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2436857600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_80_bfp.out18_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2439479040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_65_bfp.out13_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2442100480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_21_bfp.out9_411": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2444721920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_94_bfp.out9_412": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2455207680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_23_bfp.out22_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2465693440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_95_bfp.out9_413": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2476179200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_81_bfp.out18_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2478800640 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_66_bfp.out13_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2481422080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_66_bfp.out9_415": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2484043520 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_67_bfp.out9_414": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2486664960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_68_bfp.out9_416": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2489286400 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_44_bfp.out11_0_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2491907840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_96_bfp.out9_417": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2494529280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_82_bfp.out18_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2497150720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_67_bfp.out13_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2499772160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2502393600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2505015040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2510257920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.out9_418": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2515500800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_97_bfp.out9_418": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2518122240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_83_bfp.out18_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2520743680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_68_bfp.out13_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2523365120 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_22_bfp.out9_419": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2525986560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_98_bfp.out9_420": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2536472320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_24_bfp.out22_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2546958080 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_99_bfp.out9_421": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2557443840 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_84_bfp.out18_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2560065280 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_69_bfp.out13_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2562686720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_69_bfp.out9_423": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2565308160 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_70_bfp.out9_422": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2567929600 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_71_bfp.out9_424": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2570551040 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_46_bfp.out11_0_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2573172480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_100_bfp.out9_425": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2575793920 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_85_bfp.out18_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2578415360 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_70_bfp.out13_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2581036800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2583658240 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2586279680 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2591522560 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.out9_426": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2596765440 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_101_bfp.out9_426": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2599386880 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_86_bfp.out18_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2602008320 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_71_bfp.out13_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2604629760 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_23_bfp.out9_427": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2607251200 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_102_bfp.out9_428": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2617736960 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_25_bfp.out22_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2628222720 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_103_bfp.out9_429": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2638708480 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_87_bfp.out18_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2641329920 + }, + "/unet/down_blocks.2/attentions.1/Reshape_1_output_0.out2_0_20_bfp.out9_430": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2643951360 + }, + "/unet/down_blocks.2/attentions.1/Add.out_17_1_88_bfp.out18_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2646572800 + }, + "GroupNorm_16.out7_16.out6_0_12_bfp.out7_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2649194240 + }, + "NhwcConv_17_out-/unet/mid_block/resnets.0/conv1/Conv_output_0.out0_1_17_bfp.out1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2651815680 + }, + "/unet/mid_block/resnets.0/Add.out_17_1_89_bfp.out18_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2654437120 + }, + "GroupNorm_17.out7_17.out6_0_13_bfp.out7_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2657058560 + }, + "NhwcConv_18_out-/unet/mid_block/resnets.0/conv2/Conv_output_0.out0_1_18_bfp.out1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2659680000 + }, + "/unet/mid_block/resnets.0/Add_1.out_17_1_90_bfp.out18_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2662301440 + }, + "/unet/mid_block/attentions.0/norm/Add_output_0_NHWC.out6_1_4_bfp.out7_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2664922880 + }, + "/unet/mid_block/attentions.0/proj_in/Add_output_0.out2_3_104_bfp.out9_431": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2667544320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_72_bfp.out13_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2670165760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_72_bfp.out9_433": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2672787200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_73_bfp.out9_432": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2675408640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_74_bfp.out9_434": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2678030080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_48_bfp.out11_0_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2680651520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_105_bfp.out9_435": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2683272960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_91_bfp.out18_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2685894400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_73_bfp.out13_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2688515840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2691137280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2693758720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2699001600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.out9_436": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2704244480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_106_bfp.out9_436": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2706865920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_92_bfp.out18_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2709487360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_74_bfp.out13_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2712108800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_24_bfp.out9_437": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2714730240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_107_bfp.out9_438": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2725216000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_26_bfp.out22_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2735701760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_108_bfp.out9_439": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2746187520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_93_bfp.out18_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2748808960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_75_bfp.out13_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2751430400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_75_bfp.out9_441": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2754051840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_76_bfp.out9_440": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2756673280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_77_bfp.out9_442": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2759294720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_50_bfp.out11_0_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2761916160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_109_bfp.out9_443": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2764537600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_94_bfp.out18_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2767159040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_76_bfp.out13_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2769780480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2772401920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2775023360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2780266240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.out9_444": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2785509120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_110_bfp.out9_444": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2788130560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_95_bfp.out18_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2790752000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_77_bfp.out13_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2793373440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_25_bfp.out9_445": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2795994880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_111_bfp.out9_446": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2806480640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_27_bfp.out22_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2816966400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_112_bfp.out9_447": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2827452160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_96_bfp.out18_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2830073600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_78_bfp.out13_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2832695040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_78_bfp.out9_449": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2835316480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_79_bfp.out9_448": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2837937920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_80_bfp.out9_450": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2840559360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_52_bfp.out11_0_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2843180800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_113_bfp.out9_451": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2845802240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_97_bfp.out18_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2848423680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_79_bfp.out13_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2851045120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2853666560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2856288000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2861530880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.out9_452": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2866773760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_114_bfp.out9_452": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2869395200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_98_bfp.out18_98": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2872016640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_80_bfp.out13_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2874638080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_26_bfp.out9_453": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2877259520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_115_bfp.out9_454": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2887745280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_28_bfp.out22_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2898231040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_116_bfp.out9_455": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2908716800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_99_bfp.out18_99": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2911338240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_81_bfp.out13_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2913959680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_81_bfp.out9_457": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2916581120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_82_bfp.out9_456": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2919202560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_83_bfp.out9_458": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2921824000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_54_bfp.out11_0_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2924445440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_117_bfp.out9_459": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2927066880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_100_bfp.out18_100": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2929688320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_82_bfp.out13_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2932309760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2934931200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2937552640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 2942795520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.out9_460": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2948038400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_118_bfp.out9_460": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2950659840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_101_bfp.out18_101": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2953281280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_83_bfp.out13_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2955902720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_27_bfp.out9_461": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2958524160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_119_bfp.out9_462": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2969009920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_29_bfp.out22_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 2979495680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_120_bfp.out9_463": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2989981440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_102_bfp.out18_102": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2992602880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_84_bfp.out13_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2995224320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_84_bfp.out9_465": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 2997845760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_85_bfp.out9_464": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3000467200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_86_bfp.out9_466": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3003088640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_56_bfp.out11_0_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3005710080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_121_bfp.out9_467": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3008331520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_103_bfp.out18_103": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3010952960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_85_bfp.out13_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3013574400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3016195840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3018817280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3024060160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.out9_468": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3029303040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_122_bfp.out9_468": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3031924480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_104_bfp.out18_104": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3034545920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_86_bfp.out13_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3037167360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_28_bfp.out9_469": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3039788800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_123_bfp.out9_470": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3050274560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_30_bfp.out22_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3060760320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_124_bfp.out9_471": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3071246080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_105_bfp.out18_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3073867520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_87_bfp.out13_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3076488960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_87_bfp.out9_473": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3079110400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_88_bfp.out9_472": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3081731840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_89_bfp.out9_474": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3084353280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_58_bfp.out11_0_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3086974720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_125_bfp.out9_475": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3089596160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_106_bfp.out18_106": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3092217600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_88_bfp.out13_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3094839040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3097460480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3100081920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3105324800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.out9_476": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3110567680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_126_bfp.out9_476": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3113189120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_107_bfp.out18_107": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3115810560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_89_bfp.out13_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3118432000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_29_bfp.out9_477": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3121053440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_127_bfp.out9_478": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3131539200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_31_bfp.out22_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3142024960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_128_bfp.out9_479": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3152510720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_108_bfp.out18_108": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3155132160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_90_bfp.out13_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3157753600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_90_bfp.out9_481": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3160375040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_91_bfp.out9_480": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3162996480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_92_bfp.out9_482": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3165617920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_60_bfp.out11_0_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3168239360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_129_bfp.out9_483": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3170860800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_109_bfp.out18_109": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3173482240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_91_bfp.out13_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3176103680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3178725120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3181346560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3186589440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.out9_484": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3191832320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_130_bfp.out9_484": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3194453760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_110_bfp.out18_110": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3197075200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_92_bfp.out13_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3199696640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_30_bfp.out9_485": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3202318080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_131_bfp.out9_486": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3212803840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_32_bfp.out22_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3223289600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_132_bfp.out9_487": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3233775360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_111_bfp.out18_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3236396800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_93_bfp.out13_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3239018240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_93_bfp.out9_489": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3241639680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_94_bfp.out9_488": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3244261120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_95_bfp.out9_490": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3246882560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_62_bfp.out11_0_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3249504000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_133_bfp.out9_491": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3252125440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_112_bfp.out18_112": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3254746880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_94_bfp.out13_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3257368320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3259989760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3262611200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3267854080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.out9_492": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3273096960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_134_bfp.out9_492": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3275718400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_113_bfp.out18_113": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3278339840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_95_bfp.out13_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3280961280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_31_bfp.out9_493": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3283582720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_135_bfp.out9_494": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3294068480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_33_bfp.out22_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3304554240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_136_bfp.out9_495": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3315040000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_114_bfp.out18_114": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3317661440 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_96_bfp.out13_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3320282880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_96_bfp.out9_497": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3322904320 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_97_bfp.out9_496": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3325525760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_98_bfp.out9_498": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3328147200 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_64_bfp.out11_0_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3330768640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_137_bfp.out9_499": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3333390080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_115_bfp.out18_115": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3336011520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_97_bfp.out13_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3338632960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3341254400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3343875840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3349118720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.out9_500": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3354361600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_138_bfp.out9_500": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3356983040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_116_bfp.out18_116": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3359604480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_98_bfp.out13_98": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3362225920 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_32_bfp.out9_501": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3364847360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_139_bfp.out9_502": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3375333120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_34_bfp.out22_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3385818880 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_140_bfp.out9_503": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3396304640 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_117_bfp.out18_117": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3398926080 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_99_bfp.out13_99": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3401547520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_99_bfp.out9_505": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3404168960 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_100_bfp.out9_504": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3406790400 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_101_bfp.out9_506": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3409411840 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_66_bfp.out11_0_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3412033280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_141_bfp.out9_507": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3414654720 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_118_bfp.out18_118": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3417276160 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_100_bfp.out13_100": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3419897600 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3422519040 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3425140480 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3430383360 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.out9_508": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3435626240 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_142_bfp.out9_508": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3438247680 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_119_bfp.out18_119": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3440869120 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_101_bfp.out13_101": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3443490560 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_33_bfp.out9_509": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3446112000 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_143_bfp.out9_510": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3456597760 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_35_bfp.out22_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3467083520 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_144_bfp.out9_511": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3477569280 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_120_bfp.out18_120": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3480190720 + }, + "/unet/mid_block/attentions.0/Reshape_1_output_0.out2_0_21_bfp.out9_512": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3482812160 + }, + "/unet/mid_block/attentions.0/Add.out_17_1_121_bfp.out18_121": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3485433600 + }, + "GroupNorm_19.out7_19.out6_0_14_bfp.out7_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3488055040 + }, + "NhwcConv_19_out-/unet/mid_block/resnets.1/conv1/Conv_output_0.out0_1_19_bfp.out1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3490676480 + }, + "/unet/mid_block/resnets.1/Add.out_17_1_122_bfp.out18_122": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3493297920 + }, + "GroupNorm_20.out7_20.out6_0_15_bfp.out7_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3495919360 + }, + "NhwcConv_20_out-/unet/mid_block/resnets.1/conv2/Conv_output_0.out0_1_20_bfp.out1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3498540800 + }, + "/unet/mid_block/resnets.1/Add_1.out_17_1_123_bfp.out18_123": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3501162240 + }, + "/unet/up_blocks.0/Concat.out19_3_bfp.out20_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 2560 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3503783680 + }, + "GroupNorm_21.out7_21.out6_0_16_bfp.out7_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 2560 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3509026560 + }, + "NhwcConv_21_out-/unet/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_22_bfp.out1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3514269440 + }, + "/unet/up_blocks.0/resnets.0/Add.out_17_1_124_bfp.out18_124": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3516890880 + }, + "GroupNorm_22.out7_22.out6_0_17_bfp.out7_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3519512320 + }, + "NhwcConv_22_out-/unet/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_23_bfp.out1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3522133760 + }, + "NhwcConv_23_out-/unet/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_21_bfp.out1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3524755200 + }, + "/unet/up_blocks.0/resnets.0/Add_1.out_17_1_125_bfp.out18_125": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3527376640 + }, + "/unet/up_blocks.0/attentions.0/norm/Add_output_0_NHWC.out6_1_5_bfp.out7_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3529998080 + }, + "/unet/up_blocks.0/attentions.0/proj_in/Add_output_0.out2_3_145_bfp.out9_513": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3532619520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_102_bfp.out13_102": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3535240960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_102_bfp.out9_515": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3537862400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_103_bfp.out9_514": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3540483840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_104_bfp.out9_516": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3543105280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_68_bfp.out11_0_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3545726720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_146_bfp.out9_517": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3548348160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_126_bfp.out18_126": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3550969600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_103_bfp.out13_103": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3553591040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3556212480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3558833920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3564076800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.out9_518": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3569319680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_147_bfp.out9_518": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3571941120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_127_bfp.out18_127": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3574562560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_104_bfp.out13_104": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3577184000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_34_bfp.out9_519": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3579805440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_148_bfp.out9_520": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3590291200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_36_bfp.out22_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3600776960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_149_bfp.out9_521": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3611262720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_128_bfp.out18_128": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3613884160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_105_bfp.out13_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3616505600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_105_bfp.out9_523": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3619127040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_106_bfp.out9_522": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3621748480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_107_bfp.out9_524": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3624369920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_70_bfp.out11_0_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3626991360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_150_bfp.out9_525": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3629612800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_129_bfp.out18_129": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3632234240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_106_bfp.out13_106": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3634855680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3637477120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3640098560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3645341440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.out9_526": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3650584320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_151_bfp.out9_526": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3653205760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_130_bfp.out18_130": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3655827200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_107_bfp.out13_107": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3658448640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_35_bfp.out9_527": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3661070080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_152_bfp.out9_528": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3671555840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_37_bfp.out22_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3682041600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_153_bfp.out9_529": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3692527360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_131_bfp.out18_131": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3695148800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_108_bfp.out13_108": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3697770240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_108_bfp.out9_531": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3700391680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_109_bfp.out9_530": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3703013120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_110_bfp.out9_532": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3705634560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/Reshape_3_output_0.out8_72_bfp.out11_0_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3708256000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_154_bfp.out9_533": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3710877440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_output_0.out_17_1_132_bfp.out18_132": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3713498880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_109_bfp.out13_109": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3716120320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3718741760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3721363200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out8_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3726606080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.out9_534": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3731848960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_155_bfp.out9_534": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3734470400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_1_output_0.out_17_1_133_bfp.out18_133": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3737091840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_110_bfp.out13_110": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3739713280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_36_bfp.out9_535": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3742334720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_156_bfp.out9_536": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3752820480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_38_bfp.out22_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3763306240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_157_bfp.out9_537": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3773792000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/Add_2_output_0.out_17_1_134_bfp.out18_134": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3776413440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_111_bfp.out13_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3779034880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_111_bfp.out9_539": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3781656320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_112_bfp.out9_538": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3784277760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_113_bfp.out9_540": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3786899200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/Reshape_3_output_0.out8_74_bfp.out11_0_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3789520640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_158_bfp.out9_541": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3792142080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_output_0.out_17_1_135_bfp.out18_135": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3794763520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_112_bfp.out13_112": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3797384960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3800006400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3802627840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out8_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3807870720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.out9_542": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3813113600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_159_bfp.out9_542": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3815735040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_1_output_0.out_17_1_136_bfp.out18_136": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3818356480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_113_bfp.out13_113": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3820977920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_37_bfp.out9_543": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3823599360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_160_bfp.out9_544": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3834085120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_39_bfp.out22_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3844570880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_161_bfp.out9_545": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3855056640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/Add_2_output_0.out_17_1_137_bfp.out18_137": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3857678080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_114_bfp.out13_114": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3860299520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_114_bfp.out9_547": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3862920960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_115_bfp.out9_546": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3865542400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_116_bfp.out9_548": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3868163840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/Reshape_3_output_0.out8_76_bfp.out11_0_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3870785280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_162_bfp.out9_549": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3873406720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_output_0.out_17_1_138_bfp.out18_138": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3876028160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_115_bfp.out13_115": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3878649600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3881271040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3883892480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out8_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3889135360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.out9_550": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3894378240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_163_bfp.out9_550": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3896999680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_1_output_0.out_17_1_139_bfp.out18_139": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3899621120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_116_bfp.out13_116": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3902242560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_38_bfp.out9_551": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3904864000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_164_bfp.out9_552": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3915349760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_40_bfp.out22_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3925835520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_165_bfp.out9_553": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3936321280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/Add_2_output_0.out_17_1_140_bfp.out18_140": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3938942720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_117_bfp.out13_117": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3941564160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_117_bfp.out9_555": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3944185600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_118_bfp.out9_554": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3946807040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_119_bfp.out9_556": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3949428480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/Reshape_3_output_0.out8_78_bfp.out11_0_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3952049920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_166_bfp.out9_557": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3954671360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_output_0.out_17_1_141_bfp.out18_141": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3957292800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_118_bfp.out13_118": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3959914240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3962535680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3965157120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out8_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 3970400000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.out9_558": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3975642880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_167_bfp.out9_558": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3978264320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_1_output_0.out_17_1_142_bfp.out18_142": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3980885760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_119_bfp.out13_119": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 3983507200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_39_bfp.out9_559": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3986128640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_168_bfp.out9_560": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 3996614400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_41_bfp.out22_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4007100160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_169_bfp.out9_561": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4017585920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/Add_2_output_0.out_17_1_143_bfp.out18_143": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4020207360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_120_bfp.out13_120": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4022828800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_120_bfp.out9_563": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4025450240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_121_bfp.out9_562": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4028071680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_122_bfp.out9_564": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4030693120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/Reshape_3_output_0.out8_80_bfp.out11_0_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4033314560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_170_bfp.out9_565": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4035936000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_output_0.out_17_1_144_bfp.out18_144": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4038557440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_121_bfp.out13_121": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4041178880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4043800320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4046421760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out8_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4051664640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.out9_566": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4056907520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_171_bfp.out9_566": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4059528960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_1_output_0.out_17_1_145_bfp.out18_145": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4062150400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_122_bfp.out13_122": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4064771840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_40_bfp.out9_567": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4067393280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_172_bfp.out9_568": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4077879040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_42_bfp.out22_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4088364800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_173_bfp.out9_569": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4098850560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/Add_2_output_0.out_17_1_146_bfp.out18_146": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4101472000 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_123_bfp.out13_123": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4104093440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_123_bfp.out9_571": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4106714880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_124_bfp.out9_570": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4109336320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_125_bfp.out9_572": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4111957760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/Reshape_3_output_0.out8_82_bfp.out11_0_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4114579200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_174_bfp.out9_573": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4117200640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_output_0.out_17_1_147_bfp.out18_147": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4119822080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_124_bfp.out13_124": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4122443520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4125064960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4127686400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out8_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4132929280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.out9_574": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4138172160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_175_bfp.out9_574": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4140793600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_1_output_0.out_17_1_148_bfp.out18_148": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4143415040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_125_bfp.out13_125": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4146036480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_41_bfp.out9_575": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4148657920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_176_bfp.out9_576": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4159143680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_43_bfp.out22_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4169629440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_177_bfp.out9_577": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4180115200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/Add_2_output_0.out_17_1_149_bfp.out18_149": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4182736640 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_126_bfp.out13_126": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4185358080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_126_bfp.out9_579": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4187979520 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_127_bfp.out9_578": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4190600960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_128_bfp.out9_580": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4193222400 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/Reshape_3_output_0.out8_84_bfp.out11_0_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4195843840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_178_bfp.out9_581": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4198465280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_output_0.out_17_1_150_bfp.out18_150": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4201086720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_127_bfp.out13_127": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4203708160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4206329600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4208951040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out8_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4214193920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.out9_582": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4219436800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_179_bfp.out9_582": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4222058240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_1_output_0.out_17_1_151_bfp.out18_151": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4224679680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_128_bfp.out13_128": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4227301120 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_42_bfp.out9_583": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4229922560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_180_bfp.out9_584": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4240408320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_44_bfp.out22_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4250894080 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_181_bfp.out9_585": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4261379840 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/Add_2_output_0.out_17_1_152_bfp.out18_152": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4264001280 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_129_bfp.out13_129": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4266622720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_129_bfp.out9_587": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4269244160 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_130_bfp.out9_586": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4271865600 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_131_bfp.out9_588": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4274487040 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/Reshape_3_output_0.out8_86_bfp.out11_0_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4277108480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_182_bfp.out9_589": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4279729920 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_output_0.out_17_1_153_bfp.out18_153": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4282351360 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_130_bfp.out13_130": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4284972800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4287594240 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4290215680 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out8_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4295458560 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.out9_590": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4300701440 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_183_bfp.out9_590": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4303322880 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_1_output_0.out_17_1_154_bfp.out18_154": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4305944320 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_131_bfp.out13_131": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4308565760 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_43_bfp.out9_591": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4311187200 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_184_bfp.out9_592": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4321672960 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_45_bfp.out22_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4332158720 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_185_bfp.out9_593": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4342644480 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/Add_2_output_0.out_17_1_155_bfp.out18_155": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4345265920 + }, + "/unet/up_blocks.0/attentions.0/Reshape_1_output_0.out2_0_22_bfp.out9_594": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4347887360 + }, + "/unet/up_blocks.0/attentions.0/Add.out_17_1_156_bfp.out18_156": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4350508800 + }, + "/unet/up_blocks.0/Concat_1.out19_4_bfp.out20_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 2560 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4353130240 + }, + "GroupNorm_24.out7_24.out6_0_18_bfp.out7_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 2560 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4358373120 + }, + "NhwcConv_24_out-/unet/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_25_bfp.out1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4363616000 + }, + "/unet/up_blocks.0/resnets.1/Add.out_17_1_157_bfp.out18_157": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4366237440 + }, + "GroupNorm_25.out7_25.out6_0_19_bfp.out7_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4368858880 + }, + "NhwcConv_25_out-/unet/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_26_bfp.out1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4371480320 + }, + "NhwcConv_26_out-/unet/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_24_bfp.out1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4374101760 + }, + "/unet/up_blocks.0/resnets.1/Add_1.out_17_1_158_bfp.out18_158": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4376723200 + }, + "/unet/up_blocks.0/attentions.1/norm/Add_output_0_NHWC.out6_1_6_bfp.out7_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4379344640 + }, + "/unet/up_blocks.0/attentions.1/proj_in/Add_output_0.out2_3_186_bfp.out9_595": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4381966080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_132_bfp.out13_132": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4384587520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_132_bfp.out9_597": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4387208960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_133_bfp.out9_596": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4389830400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_134_bfp.out9_598": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4392451840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_88_bfp.out11_0_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4395073280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_187_bfp.out9_599": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4397694720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_159_bfp.out18_159": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4400316160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_133_bfp.out13_133": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4402937600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4405559040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4408180480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4413423360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.out9_600": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4418666240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_188_bfp.out9_600": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4421287680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_160_bfp.out18_160": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4423909120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_134_bfp.out13_134": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4426530560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_44_bfp.out9_601": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4429152000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_189_bfp.out9_602": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4439637760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_46_bfp.out22_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4450123520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_190_bfp.out9_603": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4460609280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_161_bfp.out18_161": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4463230720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_135_bfp.out13_135": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4465852160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_135_bfp.out9_605": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4468473600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_136_bfp.out9_604": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4471095040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_137_bfp.out9_606": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4473716480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_90_bfp.out11_0_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4476337920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_191_bfp.out9_607": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4478959360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_162_bfp.out18_162": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4481580800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_136_bfp.out13_136": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4484202240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4486823680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4489445120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4494688000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.out9_608": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4499930880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_192_bfp.out9_608": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4502552320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_163_bfp.out18_163": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4505173760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_137_bfp.out13_137": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4507795200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_45_bfp.out9_609": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4510416640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_193_bfp.out9_610": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4520902400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_47_bfp.out22_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4531388160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_194_bfp.out9_611": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4541873920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_164_bfp.out18_164": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4544495360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_138_bfp.out13_138": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4547116800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_138_bfp.out9_613": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4549738240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_139_bfp.out9_612": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4552359680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_140_bfp.out9_614": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4554981120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/Reshape_3_output_0.out8_92_bfp.out11_0_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4557602560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_195_bfp.out9_615": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4560224000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_output_0.out_17_1_165_bfp.out18_165": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4562845440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_139_bfp.out13_139": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4565466880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4568088320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4570709760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out8_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4575952640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.out9_616": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4581195520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_196_bfp.out9_616": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4583816960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_1_output_0.out_17_1_166_bfp.out18_166": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4586438400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_140_bfp.out13_140": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4589059840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_46_bfp.out9_617": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4591681280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_197_bfp.out9_618": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4602167040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_48_bfp.out22_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4612652800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_198_bfp.out9_619": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4623138560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/Add_2_output_0.out_17_1_167_bfp.out18_167": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4625760000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_141_bfp.out13_141": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4628381440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_141_bfp.out9_621": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4631002880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_142_bfp.out9_620": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4633624320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_143_bfp.out9_622": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4636245760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/Reshape_3_output_0.out8_94_bfp.out11_0_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4638867200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_199_bfp.out9_623": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4641488640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_output_0.out_17_1_168_bfp.out18_168": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4644110080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_142_bfp.out13_142": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4646731520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4649352960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4651974400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out8_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4657217280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.out9_624": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4662460160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_200_bfp.out9_624": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4665081600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_1_output_0.out_17_1_169_bfp.out18_169": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4667703040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_143_bfp.out13_143": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4670324480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_47_bfp.out9_625": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4672945920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_201_bfp.out9_626": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4683431680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_49_bfp.out22_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4693917440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_202_bfp.out9_627": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4704403200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/Add_2_output_0.out_17_1_170_bfp.out18_170": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4707024640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_144_bfp.out13_144": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4709646080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_144_bfp.out9_629": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4712267520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_145_bfp.out9_628": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4714888960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_146_bfp.out9_630": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4717510400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/Reshape_3_output_0.out8_96_bfp.out11_0_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4720131840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_203_bfp.out9_631": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4722753280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_output_0.out_17_1_171_bfp.out18_171": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4725374720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_145_bfp.out13_145": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4727996160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4730617600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4733239040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out8_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4738481920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.out9_632": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4743724800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_204_bfp.out9_632": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4746346240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_1_output_0.out_17_1_172_bfp.out18_172": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4748967680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_146_bfp.out13_146": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4751589120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_48_bfp.out9_633": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4754210560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_205_bfp.out9_634": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4764696320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_50_bfp.out22_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4775182080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_206_bfp.out9_635": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4785667840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/Add_2_output_0.out_17_1_173_bfp.out18_173": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4788289280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_147_bfp.out13_147": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4790910720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_147_bfp.out9_637": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4793532160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_148_bfp.out9_636": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4796153600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_149_bfp.out9_638": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4798775040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/Reshape_3_output_0.out8_98_bfp.out11_0_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4801396480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_207_bfp.out9_639": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4804017920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_output_0.out_17_1_174_bfp.out18_174": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4806639360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_148_bfp.out13_148": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4809260800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4811882240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4814503680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out8_99": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4819746560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.out9_640": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4824989440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_208_bfp.out9_640": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4827610880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_1_output_0.out_17_1_175_bfp.out18_175": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4830232320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_149_bfp.out13_149": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4832853760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_49_bfp.out9_641": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4835475200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_209_bfp.out9_642": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4845960960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_51_bfp.out22_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4856446720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_210_bfp.out9_643": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4866932480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/Add_2_output_0.out_17_1_176_bfp.out18_176": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4869553920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_150_bfp.out13_150": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4872175360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_150_bfp.out9_645": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4874796800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_151_bfp.out9_644": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4877418240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_152_bfp.out9_646": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4880039680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/Reshape_3_output_0.out8_100_bfp.out11_0_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4882661120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_211_bfp.out9_647": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4885282560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_output_0.out_17_1_177_bfp.out18_177": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4887904000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_151_bfp.out13_151": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4890525440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4893146880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4895768320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out8_101": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4901011200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.out9_648": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4906254080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_212_bfp.out9_648": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4908875520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_1_output_0.out_17_1_178_bfp.out18_178": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4911496960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_152_bfp.out13_152": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4914118400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_50_bfp.out9_649": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4916739840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_213_bfp.out9_650": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4927225600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_52_bfp.out22_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4937711360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_214_bfp.out9_651": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4948197120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/Add_2_output_0.out_17_1_179_bfp.out18_179": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4950818560 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_153_bfp.out13_153": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4953440000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_153_bfp.out9_653": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4956061440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_154_bfp.out9_652": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4958682880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_155_bfp.out9_654": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4961304320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/Reshape_3_output_0.out8_102_bfp.out11_0_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4963925760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_215_bfp.out9_655": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4966547200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_output_0.out_17_1_180_bfp.out18_180": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4969168640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_154_bfp.out13_154": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4971790080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4974411520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4977032960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out8_103": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 4982275840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.out9_656": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4987518720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_216_bfp.out9_656": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4990140160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_1_output_0.out_17_1_181_bfp.out18_181": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4992761600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_155_bfp.out13_155": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 4995383040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_51_bfp.out9_657": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 4998004480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_217_bfp.out9_658": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5008490240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_53_bfp.out22_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5018976000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_218_bfp.out9_659": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5029461760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/Add_2_output_0.out_17_1_182_bfp.out18_182": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5032083200 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_156_bfp.out13_156": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5034704640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_156_bfp.out9_661": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5037326080 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_157_bfp.out9_660": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5039947520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_158_bfp.out9_662": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5042568960 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/Reshape_3_output_0.out8_104_bfp.out11_0_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5045190400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_219_bfp.out9_663": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5047811840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_output_0.out_17_1_183_bfp.out18_183": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5050433280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_157_bfp.out13_157": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5053054720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5055676160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5058297600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out8_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5063540480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.out9_664": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5068783360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_220_bfp.out9_664": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5071404800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_1_output_0.out_17_1_184_bfp.out18_184": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5074026240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_158_bfp.out13_158": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5076647680 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_52_bfp.out9_665": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5079269120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_221_bfp.out9_666": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5089754880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_54_bfp.out22_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5100240640 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_222_bfp.out9_667": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5110726400 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/Add_2_output_0.out_17_1_185_bfp.out18_185": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5113347840 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_159_bfp.out13_159": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5115969280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_159_bfp.out9_669": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5118590720 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_160_bfp.out9_668": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5121212160 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_161_bfp.out9_670": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5123833600 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/Reshape_3_output_0.out8_106_bfp.out11_0_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5126455040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_223_bfp.out9_671": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5129076480 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_output_0.out_17_1_186_bfp.out18_186": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5131697920 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_160_bfp.out13_160": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5134319360 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5136940800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5139562240 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out8_107": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5144805120 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.out9_672": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5150048000 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_224_bfp.out9_672": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5152669440 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_1_output_0.out_17_1_187_bfp.out18_187": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5155290880 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_161_bfp.out13_161": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5157912320 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_53_bfp.out9_673": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5160533760 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_225_bfp.out9_674": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5171019520 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_55_bfp.out22_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5181505280 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_226_bfp.out9_675": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5191991040 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/Add_2_output_0.out_17_1_188_bfp.out18_188": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5194612480 + }, + "/unet/up_blocks.0/attentions.1/Reshape_1_output_0.out2_0_23_bfp.out9_676": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5197233920 + }, + "/unet/up_blocks.0/attentions.1/Add.out_17_1_189_bfp.out18_189": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5199855360 + }, + "/unet/up_blocks.0/Concat_2.out19_5_bfp.out20_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1920 + ], + "size_in_bytes": 3932160, + "op_tensor_size": 3932160, + "offset": 5202476800 + }, + "GroupNorm_27.out7_27.out6_0_20_bfp.out7_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1920 + ], + "size_in_bytes": 3932160, + "op_tensor_size": 3932160, + "offset": 5206408960 + }, + "NhwcConv_27_out-/unet/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_28_bfp.out1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5210341120 + }, + "/unet/up_blocks.0/resnets.2/Add.out_17_1_190_bfp.out18_190": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5212962560 + }, + "GroupNorm_28.out7_28.out6_0_21_bfp.out7_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5215584000 + }, + "NhwcConv_28_out-/unet/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_29_bfp.out1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5218205440 + }, + "NhwcConv_29_out-/unet/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_27_bfp.out1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5220826880 + }, + "/unet/up_blocks.0/resnets.2/Add_1.out_17_1_191_bfp.out18_191": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5223448320 + }, + "/unet/up_blocks.0/attentions.2/norm/Add_output_0_NHWC.out6_1_7_bfp.out7_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5226069760 + }, + "/unet/up_blocks.0/attentions.2/proj_in/Add_output_0.out2_3_227_bfp.out9_677": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5228691200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_162_bfp.out13_162": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5231312640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_162_bfp.out9_679": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5233934080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_163_bfp.out9_678": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5236555520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_164_bfp.out9_680": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5239176960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_108_bfp.out11_0_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5241798400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_228_bfp.out9_681": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5244419840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_192_bfp.out18_192": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5247041280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_163_bfp.out13_163": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5249662720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5252284160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5254905600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_109": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5260148480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.out9_682": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5265391360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_229_bfp.out9_682": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5268012800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_193_bfp.out18_193": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5270634240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_164_bfp.out13_164": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5273255680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_54_bfp.out9_683": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5275877120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_230_bfp.out9_684": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5286362880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_56_bfp.out22_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5296848640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_231_bfp.out9_685": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5307334400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_194_bfp.out18_194": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5309955840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_165_bfp.out13_165": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5312577280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_165_bfp.out9_687": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5315198720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_166_bfp.out9_686": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5317820160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_167_bfp.out9_688": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5320441600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_110_bfp.out11_0_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5323063040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_232_bfp.out9_689": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5325684480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_195_bfp.out18_195": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5328305920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_166_bfp.out13_166": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5330927360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5333548800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5336170240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5341413120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.out9_690": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5346656000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_233_bfp.out9_690": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5349277440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_196_bfp.out18_196": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5351898880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_167_bfp.out13_167": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5354520320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_55_bfp.out9_691": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5357141760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_234_bfp.out9_692": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5367627520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_57_bfp.out22_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5378113280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_235_bfp.out9_693": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5388599040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_197_bfp.out18_197": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5391220480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0.out12_168_bfp.out13_168": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5393841920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_key.out2_4_168_bfp.out9_695": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5396463360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_query.out2_4_169_bfp.out9_694": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5399084800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm1/LayerNormalization_output_0_bias_value.out2_4_170_bfp.out9_696": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5401706240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/Reshape_3_output_0.out8_112_bfp.out11_0_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5404327680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn1/to_out.0/Add_output_0.out2_3_236_bfp.out9_697": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5406949120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_output_0.out_17_1_198_bfp.out18_198": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5409570560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm2/LayerNormalization_output_0.out12_169_bfp.out13_169": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5412192000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5414813440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5417434880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out8_113": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5422677760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.out9_698": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5427920640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_out.0/Add_output_0.out2_3_237_bfp.out9_698": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5430542080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_1_output_0.out_17_1_199_bfp.out18_199": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5433163520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/norm3/LayerNormalization_output_0.out12_170_bfp.out13_170": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5435784960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_3_output_0.out2_2_56_bfp.out9_699": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5438406400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Slice_output_0.out2_3_238_bfp.out9_700": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5448892160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.0/Mul_4_output_0.out21_58_bfp.out22_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5459377920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/ff/net.2/Add_output_0.out2_3_239_bfp.out9_701": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5469863680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/Add_2_output_0.out_17_1_200_bfp.out18_200": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5472485120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0.out12_171_bfp.out13_171": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5475106560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_key.out2_4_171_bfp.out9_703": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5477728000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_query.out2_4_172_bfp.out9_702": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5480349440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm1/LayerNormalization_output_0_bias_value.out2_4_173_bfp.out9_704": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5482970880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/Reshape_3_output_0.out8_114_bfp.out11_0_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5485592320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn1/to_out.0/Add_output_0.out2_3_240_bfp.out9_705": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5488213760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_output_0.out_17_1_201_bfp.out18_201": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5490835200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm2/LayerNormalization_output_0.out12_172_bfp.out13_172": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5493456640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5496078080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5498699520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out8_115": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5503942400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.out9_706": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5509185280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_out.0/Add_output_0.out2_3_241_bfp.out9_706": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5511806720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_1_output_0.out_17_1_202_bfp.out18_202": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5514428160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/norm3/LayerNormalization_output_0.out12_173_bfp.out13_173": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5517049600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_3_output_0.out2_2_57_bfp.out9_707": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5519671040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Slice_output_0.out2_3_242_bfp.out9_708": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5530156800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.0/Mul_4_output_0.out21_59_bfp.out22_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5540642560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/ff/net.2/Add_output_0.out2_3_243_bfp.out9_709": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5551128320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/Add_2_output_0.out_17_1_203_bfp.out18_203": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5553749760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0.out12_174_bfp.out13_174": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5556371200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_key.out2_4_174_bfp.out9_711": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5558992640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_query.out2_4_175_bfp.out9_710": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5561614080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm1/LayerNormalization_output_0_bias_value.out2_4_176_bfp.out9_712": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5564235520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/Reshape_3_output_0.out8_116_bfp.out11_0_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5566856960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn1/to_out.0/Add_output_0.out2_3_244_bfp.out9_713": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5569478400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_output_0.out_17_1_204_bfp.out18_204": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5572099840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm2/LayerNormalization_output_0.out12_175_bfp.out13_175": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5574721280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5577342720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5579964160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out8_117": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5585207040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.out9_714": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5590449920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_out.0/Add_output_0.out2_3_245_bfp.out9_714": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5593071360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_1_output_0.out_17_1_205_bfp.out18_205": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5595692800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/norm3/LayerNormalization_output_0.out12_176_bfp.out13_176": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5598314240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_3_output_0.out2_2_58_bfp.out9_715": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5600935680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Slice_output_0.out2_3_246_bfp.out9_716": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5611421440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.0/Mul_4_output_0.out21_60_bfp.out22_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5621907200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/ff/net.2/Add_output_0.out2_3_247_bfp.out9_717": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5632392960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/Add_2_output_0.out_17_1_206_bfp.out18_206": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5635014400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0.out12_177_bfp.out13_177": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5637635840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_key.out2_4_177_bfp.out9_719": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5640257280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_query.out2_4_178_bfp.out9_718": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5642878720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm1/LayerNormalization_output_0_bias_value.out2_4_179_bfp.out9_720": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5645500160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/Reshape_3_output_0.out8_118_bfp.out11_0_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5648121600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn1/to_out.0/Add_output_0.out2_3_248_bfp.out9_721": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5650743040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_output_0.out_17_1_207_bfp.out18_207": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5653364480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm2/LayerNormalization_output_0.out12_178_bfp.out13_178": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5655985920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5658607360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5661228800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out8_119": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5666471680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.out9_722": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5671714560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_out.0/Add_output_0.out2_3_249_bfp.out9_722": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5674336000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_1_output_0.out_17_1_208_bfp.out18_208": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5676957440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/norm3/LayerNormalization_output_0.out12_179_bfp.out13_179": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5679578880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_3_output_0.out2_2_59_bfp.out9_723": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5682200320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Slice_output_0.out2_3_250_bfp.out9_724": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5692686080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.0/Mul_4_output_0.out21_61_bfp.out22_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5703171840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/ff/net.2/Add_output_0.out2_3_251_bfp.out9_725": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5713657600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/Add_2_output_0.out_17_1_209_bfp.out18_209": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5716279040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0.out12_180_bfp.out13_180": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5718900480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_key.out2_4_180_bfp.out9_727": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5721521920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_query.out2_4_181_bfp.out9_726": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5724143360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm1/LayerNormalization_output_0_bias_value.out2_4_182_bfp.out9_728": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5726764800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/Reshape_3_output_0.out8_120_bfp.out11_0_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5729386240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn1/to_out.0/Add_output_0.out2_3_252_bfp.out9_729": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5732007680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_output_0.out_17_1_210_bfp.out18_210": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5734629120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm2/LayerNormalization_output_0.out12_181_bfp.out13_181": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5737250560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5739872000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5742493440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out8_121": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5747736320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.out9_730": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5752979200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_out.0/Add_output_0.out2_3_253_bfp.out9_730": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5755600640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_1_output_0.out_17_1_211_bfp.out18_211": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5758222080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/norm3/LayerNormalization_output_0.out12_182_bfp.out13_182": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5760843520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_3_output_0.out2_2_60_bfp.out9_731": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5763464960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Slice_output_0.out2_3_254_bfp.out9_732": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5773950720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.0/Mul_4_output_0.out21_62_bfp.out22_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5784436480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/ff/net.2/Add_output_0.out2_3_255_bfp.out9_733": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5794922240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/Add_2_output_0.out_17_1_212_bfp.out18_212": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5797543680 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0.out12_183_bfp.out13_183": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5800165120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_key.out2_4_183_bfp.out9_735": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5802786560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_query.out2_4_184_bfp.out9_734": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5805408000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm1/LayerNormalization_output_0_bias_value.out2_4_185_bfp.out9_736": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5808029440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/Reshape_3_output_0.out8_122_bfp.out11_0_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5810650880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn1/to_out.0/Add_output_0.out2_3_256_bfp.out9_737": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5813272320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_output_0.out_17_1_213_bfp.out18_213": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5815893760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm2/LayerNormalization_output_0.out12_184_bfp.out13_184": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5818515200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5821136640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5823758080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out8_123": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5829000960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.out9_738": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5834243840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_out.0/Add_output_0.out2_3_257_bfp.out9_738": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5836865280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_1_output_0.out_17_1_214_bfp.out18_214": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5839486720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/norm3/LayerNormalization_output_0.out12_185_bfp.out13_185": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5842108160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_3_output_0.out2_2_61_bfp.out9_739": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5844729600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Slice_output_0.out2_3_258_bfp.out9_740": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5855215360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.0/Mul_4_output_0.out21_63_bfp.out22_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5865701120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/ff/net.2/Add_output_0.out2_3_259_bfp.out9_741": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5876186880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/Add_2_output_0.out_17_1_215_bfp.out18_215": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5878808320 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0.out12_186_bfp.out13_186": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5881429760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_key.out2_4_186_bfp.out9_743": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5884051200 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_query.out2_4_187_bfp.out9_742": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5886672640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm1/LayerNormalization_output_0_bias_value.out2_4_188_bfp.out9_744": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5889294080 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/Reshape_3_output_0.out8_124_bfp.out11_0_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5891915520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn1/to_out.0/Add_output_0.out2_3_260_bfp.out9_745": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5894536960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_output_0.out_17_1_216_bfp.out18_216": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5897158400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm2/LayerNormalization_output_0.out12_187_bfp.out13_187": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5899779840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5902401280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5905022720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out8_125": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5910265600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.out9_746": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5915508480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_out.0/Add_output_0.out2_3_261_bfp.out9_746": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5918129920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_1_output_0.out_17_1_217_bfp.out18_217": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5920751360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/norm3/LayerNormalization_output_0.out12_188_bfp.out13_188": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5923372800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_3_output_0.out2_2_62_bfp.out9_747": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5925994240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Slice_output_0.out2_3_262_bfp.out9_748": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5936480000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.0/Mul_4_output_0.out21_64_bfp.out22_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 5946965760 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/ff/net.2/Add_output_0.out2_3_263_bfp.out9_749": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5957451520 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/Add_2_output_0.out_17_1_218_bfp.out18_218": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5960072960 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0.out12_189_bfp.out13_189": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5962694400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_key.out2_4_189_bfp.out9_751": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5965315840 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_query.out2_4_190_bfp.out9_750": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5967937280 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm1/LayerNormalization_output_0_bias_value.out2_4_191_bfp.out9_752": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5970558720 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/Reshape_3_output_0.out8_126_bfp.out11_0_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5973180160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn1/to_out.0/Add_output_0.out2_3_264_bfp.out9_753": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5975801600 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_output_0.out_17_1_219_bfp.out18_219": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5978423040 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm2/LayerNormalization_output_0.out12_190_bfp.out13_190": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5981044480 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5983665920 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 20, + 1024, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5986287360 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out8_127": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 5991530240 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.out9_754": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5996773120 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_out.0/Add_output_0.out2_3_265_bfp.out9_754": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 5999394560 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_1_output_0.out_17_1_220_bfp.out18_220": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6002016000 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/norm3/LayerNormalization_output_0.out12_191_bfp.out13_191": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6004637440 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_3_output_0.out2_2_63_bfp.out9_755": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6007258880 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Slice_output_0.out2_3_266_bfp.out9_756": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6017744640 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.0/Mul_4_output_0.out21_65_bfp.out22_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 5120 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6028230400 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/ff/net.2/Add_output_0.out2_3_267_bfp.out9_757": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6038716160 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/Add_2_output_0.out_17_1_221_bfp.out18_221": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 1024, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6041337600 + }, + "/unet/up_blocks.0/attentions.2/Reshape_1_output_0.out2_0_24_bfp.out9_758": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6043959040 + }, + "/unet/up_blocks.0/attentions.2/Add.out_17_1_222_bfp.out18_222": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 2621440, + "op_tensor_size": 2621440, + "offset": 6046580480 + }, + "/unet/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_3_0_bfp.out4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1280 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6049201920 + }, + "NhwcConv_30_out-/unet/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_30_bfp.out1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1280 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6059687680 + }, + "/unet/up_blocks.1/Concat.out19_6_bfp.out20_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1920 + ], + "size_in_bytes": 15728640, + "op_tensor_size": 15728640, + "offset": 6070173440 + }, + "GroupNorm_30.out7_30.out6_0_22_bfp.out7_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1920 + ], + "size_in_bytes": 15728640, + "op_tensor_size": 15728640, + "offset": 6085902080 + }, + "NhwcConv_31_out-/unet/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_32_bfp.out1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6101630720 + }, + "/unet/up_blocks.1/resnets.0/Add.out_17_1_223_bfp.out18_223": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6106873600 + }, + "GroupNorm_31.out7_31.out6_0_23_bfp.out7_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6112116480 + }, + "NhwcConv_32_out-/unet/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_33_bfp.out1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6117359360 + }, + "NhwcConv_33_out-/unet/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_31_bfp.out1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6122602240 + }, + "/unet/up_blocks.1/resnets.0/Add_1.out_17_1_224_bfp.out18_224": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6127845120 + }, + "/unet/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out6_1_8_bfp.out7_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6133088000 + }, + "/unet/up_blocks.1/attentions.0/proj_in/Add_output_0.out2_3_268_bfp.out9_759": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6138330880 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_192_bfp.out13_192": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6143573760 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_192_bfp.out9_761": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6148816640 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_193_bfp.out9_760": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6154059520 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_194_bfp.out9_762": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6159302400 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_128_bfp.out11_0_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6164545280 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_269_bfp.out9_763": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6169788160 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_17_1_225_bfp.out18_225": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6175031040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_193_bfp.out13_193": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6180273920 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6185516800 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6190759680 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_129": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6201245440 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.out9_764": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6211731200 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_270_bfp.out9_764": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6216974080 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_17_1_226_bfp.out18_226": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6222216960 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_194_bfp.out13_194": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6227459840 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_64_bfp.out9_765": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6232702720 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_271_bfp.out9_766": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6253674240 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_66_bfp.out22_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6274645760 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_272_bfp.out9_767": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6295617280 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/Add_2_output_0.out_17_1_227_bfp.out18_227": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6300860160 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_195_bfp.out13_195": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6306103040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_195_bfp.out9_769": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6311345920 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_196_bfp.out9_768": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6316588800 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_197_bfp.out9_770": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6321831680 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/Reshape_3_output_0.out8_130_bfp.out11_0_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6327074560 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_273_bfp.out9_771": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6332317440 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_output_0.out_17_1_228_bfp.out18_228": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6337560320 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_196_bfp.out13_196": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6342803200 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6348046080 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6353288960 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out8_131": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6363774720 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.out9_772": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6374260480 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_274_bfp.out9_772": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6379503360 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_1_output_0.out_17_1_229_bfp.out18_229": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6384746240 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_197_bfp.out13_197": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6389989120 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_65_bfp.out9_773": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6395232000 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_275_bfp.out9_774": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6416203520 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_67_bfp.out22_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6437175040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_276_bfp.out9_775": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6458146560 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/Add_2_output_0.out_17_1_230_bfp.out18_230": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6463389440 + }, + "/unet/up_blocks.1/attentions.0/Reshape_1_output_0.out2_0_25_bfp.out9_776": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6468632320 + }, + "/unet/up_blocks.1/attentions.0/Add.out_17_1_231_bfp.out18_231": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6473875200 + }, + "/unet/up_blocks.1/Concat_1.out19_7_bfp.out20_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1280 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6479118080 + }, + "GroupNorm_33.out7_33.out6_0_24_bfp.out7_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 1280 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6489603840 + }, + "NhwcConv_34_out-/unet/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_35_bfp.out1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6500089600 + }, + "/unet/up_blocks.1/resnets.1/Add.out_17_1_232_bfp.out18_232": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6505332480 + }, + "GroupNorm_34.out7_34.out6_0_25_bfp.out7_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6510575360 + }, + "NhwcConv_35_out-/unet/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_36_bfp.out1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6515818240 + }, + "NhwcConv_36_out-/unet/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_34_bfp.out1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6521061120 + }, + "/unet/up_blocks.1/resnets.1/Add_1.out_17_1_233_bfp.out18_233": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6526304000 + }, + "/unet/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out6_1_9_bfp.out7_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6531546880 + }, + "/unet/up_blocks.1/attentions.1/proj_in/Add_output_0.out2_3_277_bfp.out9_777": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6536789760 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_198_bfp.out13_198": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6542032640 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_198_bfp.out9_779": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6547275520 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_199_bfp.out9_778": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6552518400 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_200_bfp.out9_780": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6557761280 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_132_bfp.out11_0_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6563004160 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_278_bfp.out9_781": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6568247040 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_17_1_234_bfp.out18_234": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6573489920 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_199_bfp.out13_199": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6578732800 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6583975680 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6589218560 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_133": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6599704320 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.out9_782": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6610190080 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_279_bfp.out9_782": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6615432960 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_17_1_235_bfp.out18_235": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6620675840 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_200_bfp.out13_200": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6625918720 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_66_bfp.out9_783": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6631161600 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_280_bfp.out9_784": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6652133120 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_68_bfp.out22_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6673104640 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_281_bfp.out9_785": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6694076160 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/Add_2_output_0.out_17_1_236_bfp.out18_236": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6699319040 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_201_bfp.out13_201": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6704561920 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_201_bfp.out9_787": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6709804800 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_202_bfp.out9_786": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6715047680 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_203_bfp.out9_788": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6720290560 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/Reshape_3_output_0.out8_134_bfp.out11_0_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6725533440 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_282_bfp.out9_789": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6730776320 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_output_0.out_17_1_237_bfp.out18_237": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6736019200 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_202_bfp.out13_202": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6741262080 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6746504960 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6751747840 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out8_135": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6762233600 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.out9_790": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6772719360 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_283_bfp.out9_790": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6777962240 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_1_output_0.out_17_1_238_bfp.out18_238": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6783205120 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_203_bfp.out13_203": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6788448000 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_67_bfp.out9_791": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6793690880 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_284_bfp.out9_792": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6814662400 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_69_bfp.out22_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 6835633920 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_285_bfp.out9_793": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6856605440 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/Add_2_output_0.out_17_1_239_bfp.out18_239": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6861848320 + }, + "/unet/up_blocks.1/attentions.1/Reshape_1_output_0.out2_0_26_bfp.out9_794": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6867091200 + }, + "/unet/up_blocks.1/attentions.1/Add.out_17_1_240_bfp.out18_240": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6872334080 + }, + "/unet/up_blocks.1/Concat_2.out19_8_bfp.out20_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 960 + ], + "size_in_bytes": 7864320, + "op_tensor_size": 7864320, + "offset": 6877576960 + }, + "GroupNorm_36.out7_36.out6_0_26_bfp.out7_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 960 + ], + "size_in_bytes": 7864320, + "op_tensor_size": 7864320, + "offset": 6885441280 + }, + "NhwcConv_37_out-/unet/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_38_bfp.out1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6893305600 + }, + "/unet/up_blocks.1/resnets.2/Add.out_17_1_241_bfp.out18_241": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6898548480 + }, + "GroupNorm_37.out7_37.out6_0_27_bfp.out7_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6903791360 + }, + "NhwcConv_38_out-/unet/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_39_bfp.out1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6909034240 + }, + "NhwcConv_39_out-/unet/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_37_bfp.out1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6914277120 + }, + "/unet/up_blocks.1/resnets.2/Add_1.out_17_1_242_bfp.out18_242": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6919520000 + }, + "/unet/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out6_1_10_bfp.out7_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6924762880 + }, + "/unet/up_blocks.1/attentions.2/proj_in/Add_output_0.out2_3_286_bfp.out9_795": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6930005760 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out12_204_bfp.out13_204": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6935248640 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_key.out2_4_204_bfp.out9_797": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6940491520 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_query.out2_4_205_bfp.out9_796": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6945734400 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0_bias_value.out2_4_206_bfp.out9_798": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6950977280 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_136_bfp.out11_0_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6956220160 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_287_bfp.out9_799": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6961463040 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_17_1_243_bfp.out18_243": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6966705920 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out12_205_bfp.out13_205": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6971948800 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 6977191680 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6982434560 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_137": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 6992920320 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.out9_800": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7003406080 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_288_bfp.out9_800": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7008648960 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_17_1_244_bfp.out18_244": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7013891840 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out12_206_bfp.out13_206": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7019134720 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_68_bfp.out9_801": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7024377600 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_289_bfp.out9_802": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7045349120 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out21_70_bfp.out22_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7066320640 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_290_bfp.out9_803": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7087292160 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/Add_2_output_0.out_17_1_245_bfp.out18_245": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7092535040 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0.out12_207_bfp.out13_207": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7097777920 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_key.out2_4_207_bfp.out9_805": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7103020800 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_query.out2_4_208_bfp.out9_804": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7108263680 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm1/LayerNormalization_output_0_bias_value.out2_4_209_bfp.out9_806": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7113506560 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/Reshape_3_output_0.out8_138_bfp.out11_0_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7118749440 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn1/to_out.0/Add_output_0.out2_3_291_bfp.out9_807": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7123992320 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_output_0.out_17_1_246_bfp.out18_246": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7129235200 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm2/LayerNormalization_output_0.out12_208_bfp.out13_208": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7134478080 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7139720960 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 10, + 4096, + 64 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7144963840 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out8_139": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7155449600 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.out9_808": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7165935360 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_out.0/Add_output_0.out2_3_292_bfp.out9_808": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7171178240 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_1_output_0.out_17_1_247_bfp.out18_247": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7176421120 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/norm3/LayerNormalization_output_0.out12_209_bfp.out13_209": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7181664000 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_3_output_0.out2_2_69_bfp.out9_809": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7186906880 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Slice_output_0.out2_3_293_bfp.out9_810": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7207878400 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.0/Mul_4_output_0.out21_71_bfp.out22_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 2560 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7228849920 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/ff/net.2/Add_output_0.out2_3_294_bfp.out9_811": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7249821440 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/Add_2_output_0.out_17_1_248_bfp.out18_248": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 4096, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7255064320 + }, + "/unet/up_blocks.1/attentions.2/Reshape_1_output_0.out2_0_27_bfp.out9_812": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7260307200 + }, + "/unet/up_blocks.1/attentions.2/Add.out_17_1_249_bfp.out18_249": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 5242880, + "op_tensor_size": 5242880, + "offset": 7265550080 + }, + "/unet/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_3_1_bfp.out4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7270792960 + }, + "NhwcConv_40_out-/unet/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_40_bfp.out1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7291764480 + }, + "/unet/up_blocks.2/Concat.out19_9_bfp.out20_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 960 + ], + "size_in_bytes": 31457280, + "op_tensor_size": 31457280, + "offset": 7312736000 + }, + "GroupNorm_39.out7_39.out6_0_28_bfp.out7_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 960 + ], + "size_in_bytes": 31457280, + "op_tensor_size": 31457280, + "offset": 7344193280 + }, + "NhwcConv_41_out-/unet/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_42_bfp.out1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7375650560 + }, + "/unet/up_blocks.2/resnets.0/Add.out_17_1_250_bfp.out18_250": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7386136320 + }, + "GroupNorm_40.out7_40.out6_0_29_bfp.out7_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7396622080 + }, + "NhwcConv_42_out-/unet/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_43_bfp.out1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7407107840 + }, + "NhwcConv_43_out-/unet/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_41_bfp.out1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7417593600 + }, + "/unet/up_blocks.2/resnets.0/Add_1.out_17_1_251_bfp.out18_251": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7428079360 + }, + "/unet/up_blocks.2/Concat_1.out19_10_bfp.out20_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7438565120 + }, + "GroupNorm_41.out7_41.out6_0_30_bfp.out7_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7459536640 + }, + "NhwcConv_44_out-/unet/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_45_bfp.out1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7480508160 + }, + "/unet/up_blocks.2/resnets.1/Add.out_17_1_252_bfp.out18_252": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7490993920 + }, + "GroupNorm_42.out7_42.out6_0_31_bfp.out7_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7501479680 + }, + "NhwcConv_45_out-/unet/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_46_bfp.out1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7511965440 + }, + "NhwcConv_46_out-/unet/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_44_bfp.out1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7522451200 + }, + "/unet/up_blocks.2/resnets.1/Add_1.out_17_1_253_bfp.out18_253": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7532936960 + }, + "/unet/up_blocks.2/Concat_2.out19_11_bfp.out20_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7543422720 + }, + "GroupNorm_43.out7_43.out6_0_32_bfp.out7_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 640 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7564394240 + }, + "NhwcConv_47_out-/unet/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_48_bfp.out1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7585365760 + }, + "/unet/up_blocks.2/resnets.2/Add.out_17_1_254_bfp.out18_254": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7595851520 + }, + "GroupNorm_44.out7_44.out6_0_33_bfp.out7_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7606337280 + }, + "NhwcConv_48_out-/unet/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_49_bfp.out1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7616823040 + }, + "NhwcConv_49_out-/unet/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_47_bfp.out1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7627308800 + }, + "/unet/up_blocks.2/resnets.2/Add_1.out_17_1_255_bfp.out18_255": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7637794560 + }, + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 10485760, + "op_tensor_size": 10485760, + "offset": 7648280320 + }, + "GroupNorm_45.out7_45.out6_0_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 128, + 128, + 320 + ], + "size_in_bytes": 20971520, + "op_tensor_size": 20971520, + "offset": 7658766080 + }, + "NhwcConv_0_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 31040 + ], + "size_in_bytes": 31040, + "op_tensor_size": 31040, + "offset": 0, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_0.const", + "file_size": 31040 + }, + "NhwcConv_0_out-/unet/conv_in/Conv_output_0.out6_0_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 31040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1.const", + "file_size": 128 + }, + "GroupNorm_0_wts_6_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 31168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_2.const", + "file_size": 1280 + }, + "unet.add_embedding.linear_1.weight_11_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "op_tensor_size": 4096000, + "offset": 32448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_3.const", + "file_size": 4096000 + }, + "/unet/add_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 4128448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_4.const", + "file_size": 256 + }, + "unet.add_embedding.linear_2.weight_11_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 4128704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_5.const", + "file_size": 1868800 + }, + "onnx::MatMul_16751": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 5997504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_6.const", + "file_size": 1515520 + }, + "onnx::MatMul_16752": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 7513024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_7.const", + "file_size": 1515520 + }, + "onnx::MatMul_16775": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 9028544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_8.const", + "file_size": 1515520 + }, + "onnx::MatMul_16776": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 10544064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_9.const", + "file_size": 1515520 + }, + "onnx::MatMul_16807": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 12059584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_10.const", + "file_size": 1515520 + }, + "onnx::MatMul_16808": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 13575104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_11.const", + "file_size": 1515520 + }, + "onnx::MatMul_16831": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 15090624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_12.const", + "file_size": 1515520 + }, + "onnx::MatMul_16832": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 16606144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_13.const", + "file_size": 1515520 + }, + "onnx::MatMul_16863": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 18121664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_14.const", + "file_size": 2990080 + }, + "onnx::MatMul_16864": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 21111744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_15.const", + "file_size": 2990080 + }, + "onnx::MatMul_16887": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 24101824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_16.const", + "file_size": 2990080 + }, + "onnx::MatMul_16888": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 27091904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_17.const", + "file_size": 2990080 + }, + "onnx::MatMul_16911": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 30081984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_18.const", + "file_size": 2990080 + }, + "onnx::MatMul_16912": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 33072064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_19.const", + "file_size": 2990080 + }, + "onnx::MatMul_16935": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 36062144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_20.const", + "file_size": 2990080 + }, + "onnx::MatMul_16936": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 39052224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_21.const", + "file_size": 2990080 + }, + "onnx::MatMul_16959": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 42042304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_22.const", + "file_size": 2990080 + }, + "onnx::MatMul_16960": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 45032384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_23.const", + "file_size": 2990080 + }, + "onnx::MatMul_16983": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 48022464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_24.const", + "file_size": 2990080 + }, + "onnx::MatMul_16984": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 51012544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_25.const", + "file_size": 2990080 + }, + "onnx::MatMul_17007": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 54002624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_26.const", + "file_size": 2990080 + }, + "onnx::MatMul_17008": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 56992704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_27.const", + "file_size": 2990080 + }, + "onnx::MatMul_17031": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 59982784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_28.const", + "file_size": 2990080 + }, + "onnx::MatMul_17032": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 62972864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_29.const", + "file_size": 2990080 + }, + "onnx::MatMul_17055": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 65962944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_30.const", + "file_size": 2990080 + }, + "onnx::MatMul_17056": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 68953024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_31.const", + "file_size": 2990080 + }, + "onnx::MatMul_17079": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 71943104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_32.const", + "file_size": 2990080 + }, + "onnx::MatMul_17080": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 74933184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_33.const", + "file_size": 2990080 + }, + "onnx::MatMul_17111": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 77923264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_34.const", + "file_size": 2990080 + }, + "onnx::MatMul_17112": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 80913344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_35.const", + "file_size": 2990080 + }, + "onnx::MatMul_17135": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 83903424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_36.const", + "file_size": 2990080 + }, + "onnx::MatMul_17136": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 86893504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_37.const", + "file_size": 2990080 + }, + "onnx::MatMul_17159": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 89883584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_38.const", + "file_size": 2990080 + }, + "onnx::MatMul_17160": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 92873664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_39.const", + "file_size": 2990080 + }, + "onnx::MatMul_17183": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 95863744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_40.const", + "file_size": 2990080 + }, + "onnx::MatMul_17184": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 98853824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_41.const", + "file_size": 2990080 + }, + "onnx::MatMul_17207": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 101843904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_42.const", + "file_size": 2990080 + }, + "onnx::MatMul_17208": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 104833984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_43.const", + "file_size": 2990080 + }, + "onnx::MatMul_17231": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 107824064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_44.const", + "file_size": 2990080 + }, + "onnx::MatMul_17232": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 110814144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_45.const", + "file_size": 2990080 + }, + "onnx::MatMul_17255": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 113804224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_46.const", + "file_size": 2990080 + }, + "onnx::MatMul_17256": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 116794304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_47.const", + "file_size": 2990080 + }, + "onnx::MatMul_17279": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 119784384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_48.const", + "file_size": 2990080 + }, + "onnx::MatMul_17280": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 122774464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_49.const", + "file_size": 2990080 + }, + "onnx::MatMul_17303": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 125764544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_50.const", + "file_size": 2990080 + }, + "onnx::MatMul_17304": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 128754624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_51.const", + "file_size": 2990080 + }, + "onnx::MatMul_17327": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 131744704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_52.const", + "file_size": 2990080 + }, + "onnx::MatMul_17328": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 134734784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_53.const", + "file_size": 2990080 + }, + "onnx::MatMul_17359": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 137724864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_54.const", + "file_size": 2990080 + }, + "onnx::MatMul_17360": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 140714944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_55.const", + "file_size": 2990080 + }, + "onnx::MatMul_17383": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 143705024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_56.const", + "file_size": 2990080 + }, + "onnx::MatMul_17384": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 146695104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_57.const", + "file_size": 2990080 + }, + "onnx::MatMul_17407": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 149685184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_58.const", + "file_size": 2990080 + }, + "onnx::MatMul_17408": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 152675264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_59.const", + "file_size": 2990080 + }, + "onnx::MatMul_17431": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 155665344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_60.const", + "file_size": 2990080 + }, + "onnx::MatMul_17432": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 158655424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_61.const", + "file_size": 2990080 + }, + "onnx::MatMul_17455": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 161645504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_62.const", + "file_size": 2990080 + }, + "onnx::MatMul_17456": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 164635584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_63.const", + "file_size": 2990080 + }, + "onnx::MatMul_17479": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 167625664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_64.const", + "file_size": 2990080 + }, + "onnx::MatMul_17480": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 170615744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_65.const", + "file_size": 2990080 + }, + "onnx::MatMul_17503": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 173605824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_66.const", + "file_size": 2990080 + }, + "onnx::MatMul_17504": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 176595904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_67.const", + "file_size": 2990080 + }, + "onnx::MatMul_17527": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 179585984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_68.const", + "file_size": 2990080 + }, + "onnx::MatMul_17528": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 182576064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_69.const", + "file_size": 2990080 + }, + "onnx::MatMul_17551": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 185566144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_70.const", + "file_size": 2990080 + }, + "onnx::MatMul_17552": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 188556224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_71.const", + "file_size": 2990080 + }, + "onnx::MatMul_17575": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 191546304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_72.const", + "file_size": 2990080 + }, + "onnx::MatMul_17576": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 194536384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_73.const", + "file_size": 2990080 + }, + "unet.time_embedding.linear_1.weight_11_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 197526464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_74.const", + "file_size": 471040 + }, + "/unet/time_embedding/act/Mul/QuickGeluFusion/_sigmoid.weights14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 197997504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_75.const", + "file_size": 256 + }, + "unet.time_embedding.linear_2.weight_11_1_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 197997760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_76.const", + "file_size": 1868800 + }, + "/unet/down_blocks.0/resnets.0/act_1/Mul/QuickGeluFusion/_sigmoid.weights14_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 199866560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_77.const", + "file_size": 256 + }, + "unet.down_blocks.0.resnets.0.time_emb_proj.weight_11_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "op_tensor_size": 467200, + "offset": 199866816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_78.const", + "file_size": 467200 + }, + "NhwcConv_1_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 200334016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_79.const", + "file_size": 1057280 + }, + "GroupNorm_1_wts_6_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 201391296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_80.const", + "file_size": 1280 + }, + "NhwcConv_2_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 201392576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_81.const", + "file_size": 1057280 + }, + "GroupNorm_2_wts_6_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 202449856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_82.const", + "file_size": 1280 + }, + "unet.down_blocks.0.resnets.1.time_emb_proj.weight_11_1_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "op_tensor_size": 467200, + "offset": 202451136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_83.const", + "file_size": 467200 + }, + "NhwcConv_3_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 202918336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_84.const", + "file_size": 1057280 + }, + "GroupNorm_3_wts_6_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 203975616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_85.const", + "file_size": 1280 + }, + "NhwcConv_4_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 203976896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_86.const", + "file_size": 1057280 + }, + "NhwcConv_5_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1062400 + ], + "size_in_bytes": 1062400, + "op_tensor_size": 1062400, + "offset": 205034176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_87.const", + "file_size": 1062400 + }, + "GroupNorm_4_wts_6_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 206096576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_88.const", + "file_size": 1280 + }, + "unet.down_blocks.1.resnets.0.time_emb_proj.weight_11_1_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "op_tensor_size": 934400, + "offset": 206097856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_89.const", + "file_size": 934400 + }, + "NhwcConv_6_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2114560 + ], + "size_in_bytes": 2114560, + "op_tensor_size": 2114560, + "offset": 207032256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_90.const", + "file_size": 2114560 + }, + "GroupNorm_5_wts_6_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 209146816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_91.const", + "file_size": 2560 + }, + "unet.down_blocks.1.resnets.1.time_emb_proj.weight_11_1_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "op_tensor_size": 934400, + "offset": 209149376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_92.const", + "file_size": 934400 + }, + "unet.down_blocks.2.resnets.0.time_emb_proj.weight_11_1_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 210083776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_93.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.resnets.1.time_emb_proj.weight_11_1_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 211952576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_94.const", + "file_size": 1868800 + }, + "unet.mid_block.resnets.0.time_emb_proj.weight_11_1_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 213821376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_95.const", + "file_size": 1868800 + }, + "unet.mid_block.resnets.1.time_emb_proj.weight_11_1_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 215690176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_96.const", + "file_size": 1868800 + }, + "onnx::MatMul_17611": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 217558976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_97.const", + "file_size": 2990080 + }, + "onnx::MatMul_17612": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 220549056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_98.const", + "file_size": 2990080 + }, + "onnx::MatMul_17635": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 223539136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_99.const", + "file_size": 2990080 + }, + "onnx::MatMul_17636": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 226529216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_100.const", + "file_size": 2990080 + }, + "onnx::MatMul_17659": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 229519296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_101.const", + "file_size": 2990080 + }, + "onnx::MatMul_17660": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 232509376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_102.const", + "file_size": 2990080 + }, + "onnx::MatMul_17683": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 235499456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_103.const", + "file_size": 2990080 + }, + "onnx::MatMul_17684": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 238489536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_104.const", + "file_size": 2990080 + }, + "onnx::MatMul_17707": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 241479616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_105.const", + "file_size": 2990080 + }, + "onnx::MatMul_17708": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 244469696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_106.const", + "file_size": 2990080 + }, + "onnx::MatMul_17731": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 247459776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_107.const", + "file_size": 2990080 + }, + "onnx::MatMul_17732": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 250449856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_108.const", + "file_size": 2990080 + }, + "onnx::MatMul_17755": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 253439936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_109.const", + "file_size": 2990080 + }, + "onnx::MatMul_17756": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 256430016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_110.const", + "file_size": 2990080 + }, + "onnx::MatMul_17779": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 259420096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_111.const", + "file_size": 2990080 + }, + "onnx::MatMul_17780": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 262410176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_112.const", + "file_size": 2990080 + }, + "onnx::MatMul_17803": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 265400256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_113.const", + "file_size": 2990080 + }, + "onnx::MatMul_17804": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 268390336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_114.const", + "file_size": 2990080 + }, + "onnx::MatMul_17827": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 271380416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_115.const", + "file_size": 2990080 + }, + "onnx::MatMul_17828": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 274370496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_116.const", + "file_size": 2990080 + }, + "onnx::MatMul_17859": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 277360576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_117.const", + "file_size": 2990080 + }, + "onnx::MatMul_17860": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 280350656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_118.const", + "file_size": 2990080 + }, + "onnx::MatMul_17883": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 283340736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_119.const", + "file_size": 2990080 + }, + "onnx::MatMul_17884": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 286330816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_120.const", + "file_size": 2990080 + }, + "onnx::MatMul_17907": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 289320896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_121.const", + "file_size": 2990080 + }, + "onnx::MatMul_17908": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 292310976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_122.const", + "file_size": 2990080 + }, + "onnx::MatMul_17931": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 295301056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_123.const", + "file_size": 2990080 + }, + "onnx::MatMul_17932": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 298291136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_124.const", + "file_size": 2990080 + }, + "onnx::MatMul_17955": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 301281216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_125.const", + "file_size": 2990080 + }, + "onnx::MatMul_17956": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 304271296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_126.const", + "file_size": 2990080 + }, + "onnx::MatMul_17979": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 307261376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_127.const", + "file_size": 2990080 + }, + "onnx::MatMul_17980": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 310251456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_128.const", + "file_size": 2990080 + }, + "onnx::MatMul_18003": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 313241536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_129.const", + "file_size": 2990080 + }, + "onnx::MatMul_18004": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 316231616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_130.const", + "file_size": 2990080 + }, + "onnx::MatMul_18027": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 319221696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_131.const", + "file_size": 2990080 + }, + "onnx::MatMul_18028": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 322211776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_132.const", + "file_size": 2990080 + }, + "onnx::MatMul_18051": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 325201856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_133.const", + "file_size": 2990080 + }, + "onnx::MatMul_18052": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 328191936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_134.const", + "file_size": 2990080 + }, + "onnx::MatMul_18075": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 331182016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_135.const", + "file_size": 2990080 + }, + "onnx::MatMul_18076": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 334172096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_136.const", + "file_size": 2990080 + }, + "onnx::MatMul_18107": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 337162176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_137.const", + "file_size": 2990080 + }, + "onnx::MatMul_18108": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 340152256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_138.const", + "file_size": 2990080 + }, + "onnx::MatMul_18131": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 343142336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_139.const", + "file_size": 2990080 + }, + "onnx::MatMul_18132": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 346132416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_140.const", + "file_size": 2990080 + }, + "onnx::MatMul_18155": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 349122496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_141.const", + "file_size": 2990080 + }, + "onnx::MatMul_18156": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 352112576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_142.const", + "file_size": 2990080 + }, + "onnx::MatMul_18179": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 355102656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_143.const", + "file_size": 2990080 + }, + "onnx::MatMul_18180": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 358092736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_144.const", + "file_size": 2990080 + }, + "onnx::MatMul_18203": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 361082816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_145.const", + "file_size": 2990080 + }, + "onnx::MatMul_18204": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 364072896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_146.const", + "file_size": 2990080 + }, + "onnx::MatMul_18227": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 367062976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_147.const", + "file_size": 2990080 + }, + "onnx::MatMul_18228": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 370053056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_148.const", + "file_size": 2990080 + }, + "onnx::MatMul_18251": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 373043136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_149.const", + "file_size": 2990080 + }, + "onnx::MatMul_18252": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 376033216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_150.const", + "file_size": 2990080 + }, + "onnx::MatMul_18275": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 379023296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_151.const", + "file_size": 2990080 + }, + "onnx::MatMul_18276": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 382013376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_152.const", + "file_size": 2990080 + }, + "onnx::MatMul_18299": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 385003456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_153.const", + "file_size": 2990080 + }, + "onnx::MatMul_18300": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 387993536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_154.const", + "file_size": 2990080 + }, + "onnx::MatMul_18323": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 390983616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_155.const", + "file_size": 2990080 + }, + "onnx::MatMul_18324": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2990080 + ], + "size_in_bytes": 2990080, + "op_tensor_size": 2990080, + "offset": 393973696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_156.const", + "file_size": 2990080 + }, + "unet.up_blocks.0.resnets.0.time_emb_proj.weight_11_1_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 396963776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_157.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.resnets.1.time_emb_proj.weight_11_1_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 398832576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_158.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.resnets.2.time_emb_proj.weight_11_1_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 400701376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_159.const", + "file_size": 1868800 + }, + "onnx::MatMul_18356": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 402570176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_160.const", + "file_size": 1515520 + }, + "onnx::MatMul_18357": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 404085696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_161.const", + "file_size": 1515520 + }, + "onnx::MatMul_18380": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 405601216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_162.const", + "file_size": 1515520 + }, + "onnx::MatMul_18381": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 407116736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_163.const", + "file_size": 1515520 + }, + "onnx::MatMul_18412": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 408632256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_164.const", + "file_size": 1515520 + }, + "onnx::MatMul_18413": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 410147776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_165.const", + "file_size": 1515520 + }, + "onnx::MatMul_18436": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 411663296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_166.const", + "file_size": 1515520 + }, + "onnx::MatMul_18437": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 413178816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_167.const", + "file_size": 1515520 + }, + "onnx::MatMul_18468": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 414694336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_168.const", + "file_size": 1515520 + }, + "onnx::MatMul_18469": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 416209856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_169.const", + "file_size": 1515520 + }, + "onnx::MatMul_18492": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 417725376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_170.const", + "file_size": 1515520 + }, + "onnx::MatMul_18493": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1515520 + ], + "size_in_bytes": 1515520, + "op_tensor_size": 1515520, + "offset": 419240896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_171.const", + "file_size": 1515520 + }, + "unet.up_blocks.1.resnets.0.time_emb_proj.weight_11_1_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "op_tensor_size": 934400, + "offset": 420756416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_172.const", + "file_size": 934400 + }, + "unet.up_blocks.1.resnets.1.time_emb_proj.weight_11_1_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "op_tensor_size": 934400, + "offset": 421690816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_173.const", + "file_size": 934400 + }, + "unet.up_blocks.1.resnets.2.time_emb_proj.weight_11_1_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "op_tensor_size": 934400, + "offset": 422625216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_174.const", + "file_size": 934400 + }, + "unet.up_blocks.2.resnets.0.time_emb_proj.weight_11_1_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "op_tensor_size": 467200, + "offset": 423559616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_175.const", + "file_size": 467200 + }, + "unet.up_blocks.2.resnets.1.time_emb_proj.weight_11_1_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "op_tensor_size": 467200, + "offset": 424026816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_176.const", + "file_size": 467200 + }, + "unet.up_blocks.2.resnets.2.time_emb_proj.weight_11_1_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "op_tensor_size": 467200, + "offset": 424494016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_177.const", + "file_size": 467200 + }, + "NhwcConv_7_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 424961216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_178.const", + "file_size": 4311040 + }, + "NhwcConv_8_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 250880 + ], + "size_in_bytes": 250880, + "op_tensor_size": 250880, + "offset": 429272256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_179.const", + "file_size": 250880 + }, + "GroupNorm_6_wts_6_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 429523136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_180.const", + "file_size": 2560 + }, + "onnx::MatMul_16738_11_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 429525696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_181.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 429996736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_182.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 429998016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_183.const", + "file_size": 1280 + }, + "Attention_0_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 429999296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_184.const", + "file_size": 471040 + }, + "Attention_0_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 430470336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_185.const", + "file_size": 471040 + }, + "Attention_0_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 430941376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_186.const", + "file_size": 471040 + }, + "onnx::MatMul_16749_11_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 431412416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_187.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 431883456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_188.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 431884736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_189.const", + "file_size": 1280 + }, + "onnx::MatMul_16750": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 431886016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_190.const", + "file_size": 471040 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_140_bfp.out9_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 432357056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_191.const", + "file_size": 128 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_6_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 432357184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_192.const", + "file_size": 128 + }, + "onnx::MatMul_16760_11_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 432357312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_193.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 432828352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_194.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 432829632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_195.const", + "file_size": 1280 + }, + "onnx::MatMul_16761_11_0_3_12_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 432830912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_196.const", + "file_size": 1868800 + }, + "onnx::MatMul_16761_11_0_3_12_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 434699712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_197.const", + "file_size": 1868800 + }, + "onnx::MatMul_16762_11_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 436568512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_198.const", + "file_size": 1884160 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 438452672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_199.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 438453952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_200.const", + "file_size": 1280 + }, + "Attention_1_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 438455232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_201.const", + "file_size": 471040 + }, + "Attention_1_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 438926272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_202.const", + "file_size": 471040 + }, + "Attention_1_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 439397312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_203.const", + "file_size": 471040 + }, + "onnx::MatMul_16773_11_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 439868352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_204.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 440339392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_205.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 440340672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_206.const", + "file_size": 1280 + }, + "onnx::MatMul_16774": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 440341952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_207.const", + "file_size": 471040 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_141_bfp.out9_3_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 440812992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_208.const", + "file_size": 128 + }, + "/unet/down_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_10_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 440813120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_209.const", + "file_size": 128 + }, + "onnx::MatMul_16784_11_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 440813248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_210.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 441284288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_211.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.0.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 441285568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_212.const", + "file_size": 1280 + }, + "onnx::MatMul_16785_11_0_7_12_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 441286848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_213.const", + "file_size": 1868800 + }, + "onnx::MatMul_16785_11_0_7_12_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 443155648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_214.const", + "file_size": 1868800 + }, + "onnx::MatMul_16786_11_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 445024448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_215.const", + "file_size": 1884160 + }, + "onnx::MatMul_16787_11_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 446908608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_216.const", + "file_size": 471040 + }, + "GroupNorm_7_wts_6_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 447379648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_217.const", + "file_size": 2560 + }, + "NhwcConv_9_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 447382208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_218.const", + "file_size": 4311040 + }, + "GroupNorm_8_wts_6_0_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 451693248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_219.const", + "file_size": 2560 + }, + "NhwcConv_10_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 451695808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_220.const", + "file_size": 4311040 + }, + "GroupNorm_9_wts_6_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 456006848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_221.const", + "file_size": 2560 + }, + "onnx::MatMul_16794_11_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 456009408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_222.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 456480448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_223.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 456481728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_224.const", + "file_size": 1280 + }, + "Attention_2_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 456483008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_225.const", + "file_size": 471040 + }, + "Attention_2_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 456954048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_226.const", + "file_size": 471040 + }, + "Attention_2_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 457425088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_227.const", + "file_size": 471040 + }, + "onnx::MatMul_16805_11_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 457896128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_228.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 458367168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_229.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 458368448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_230.const", + "file_size": 1280 + }, + "onnx::MatMul_16806": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 458369728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_231.const", + "file_size": 471040 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_142_bfp.out9_6_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 458840768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_232.const", + "file_size": 128 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_15_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 458840896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_233.const", + "file_size": 128 + }, + "onnx::MatMul_16816_11_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 458841024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_234.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 459312064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_235.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 459313344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_236.const", + "file_size": 1280 + }, + "onnx::MatMul_16817_11_0_13_12_2_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 459314624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_237.const", + "file_size": 1868800 + }, + "onnx::MatMul_16817_11_0_13_12_2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 461183424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_238.const", + "file_size": 1868800 + }, + "onnx::MatMul_16818_11_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 463052224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_239.const", + "file_size": 1884160 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 464936384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_240.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 464937664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_241.const", + "file_size": 1280 + }, + "Attention_3_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 464938944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_242.const", + "file_size": 471040 + }, + "Attention_3_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 465409984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_243.const", + "file_size": 471040 + }, + "Attention_3_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 465881024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_244.const", + "file_size": 471040 + }, + "onnx::MatMul_16829_11_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 466352064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_245.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 466823104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_246.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 466824384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_247.const", + "file_size": 1280 + }, + "onnx::MatMul_16830": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 466825664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_248.const", + "file_size": 471040 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_143_bfp.out9_9_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 467296704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_249.const", + "file_size": 128 + }, + "/unet/down_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_19_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 467296832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_250.const", + "file_size": 128 + }, + "onnx::MatMul_16840_11_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 467296960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_251.const", + "file_size": 471040 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 467768000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_252.const", + "file_size": 1280 + }, + "unet.down_blocks.1.attentions.1.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 467769280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_253.const", + "file_size": 1280 + }, + "onnx::MatMul_16841_11_0_17_12_3_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 467770560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_254.const", + "file_size": 1868800 + }, + "onnx::MatMul_16841_11_0_17_12_3_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 469639360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_255.const", + "file_size": 1868800 + }, + "onnx::MatMul_16842_11_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 471508160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_256.const", + "file_size": 1884160 + }, + "onnx::MatMul_16843_11_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 473392320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_257.const", + "file_size": 471040 + }, + "NhwcConv_11_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4249600 + ], + "size_in_bytes": 4249600, + "op_tensor_size": 4249600, + "offset": 473863360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_258.const", + "file_size": 4249600 + }, + "GroupNorm_10_wts_6_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 478112960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_259.const", + "file_size": 2560 + }, + "NhwcConv_12_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 8458240 + ], + "size_in_bytes": 8458240, + "op_tensor_size": 8458240, + "offset": 478115520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_260.const", + "file_size": 8458240 + }, + "GroupNorm_11_wts_6_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 486573760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_261.const", + "file_size": 5120 + }, + "NhwcConv_13_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 486578880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_262.const", + "file_size": 17244160 + }, + "NhwcConv_14_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1003520 + ], + "size_in_bytes": 1003520, + "op_tensor_size": 1003520, + "offset": 503823040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_263.const", + "file_size": 1003520 + }, + "GroupNorm_12_wts_6_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 504826560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_264.const", + "file_size": 5120 + }, + "onnx::MatMul_16850_11_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 504831680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_265.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 506700480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_266.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 506703040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_267.const", + "file_size": 2560 + }, + "Attention_4_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 506705600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_268.const", + "file_size": 1868800 + }, + "Attention_4_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 508574400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_269.const", + "file_size": 1868800 + }, + "Attention_4_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 510443200, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_270.const", + "file_size": 1868800 + }, + "onnx::MatMul_16861_11_0_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 512312000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_271.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 514180800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_272.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 514183360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_273.const", + "file_size": 2560 + }, + "onnx::MatMul_16862": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 514185920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_274.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_144_bfp.out9_12_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 516054720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_275.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_24_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 516054848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_276.const", + "file_size": 128 + }, + "onnx::MatMul_16872_11_0_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 516054976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_277.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 517923776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_278.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 517926336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_279.const", + "file_size": 2560 + }, + "onnx::MatMul_16873_11_0_23_12_4_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 517928896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_280.const", + "file_size": 7475200 + }, + "onnx::MatMul_16873_11_0_23_12_4_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 525404096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_281.const", + "file_size": 7475200 + }, + "onnx::MatMul_16874_11_0_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 532879296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_282.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 540354496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_283.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 540357056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_284.const", + "file_size": 2560 + }, + "Attention_5_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 540359616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_285.const", + "file_size": 1868800 + }, + "Attention_5_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 542228416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_286.const", + "file_size": 1868800 + }, + "Attention_5_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 544097216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_287.const", + "file_size": 1868800 + }, + "onnx::MatMul_16885_11_0_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 545966016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_288.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 547834816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_289.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 547837376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_290.const", + "file_size": 2560 + }, + "onnx::MatMul_16886": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 547839936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_291.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_145_bfp.out9_15_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 549708736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_292.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_28_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 549708864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_293.const", + "file_size": 128 + }, + "onnx::MatMul_16896_11_0_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 549708992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_294.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 551577792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_295.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 551580352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_296.const", + "file_size": 2560 + }, + "onnx::MatMul_16897_11_0_27_12_5_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 551582912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_297.const", + "file_size": 7475200 + }, + "onnx::MatMul_16897_11_0_27_12_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 559058112, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_298.const", + "file_size": 7475200 + }, + "onnx::MatMul_16898_11_0_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 566533312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_299.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 574008512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_300.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 574011072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_301.const", + "file_size": 2560 + }, + "Attention_6_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 574013632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_302.const", + "file_size": 1868800 + }, + "Attention_6_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 575882432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_303.const", + "file_size": 1868800 + }, + "Attention_6_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 577751232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_304.const", + "file_size": 1868800 + }, + "onnx::MatMul_16909_11_0_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 579620032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_305.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 581488832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_306.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 581491392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_307.const", + "file_size": 2560 + }, + "onnx::MatMul_16910": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 581493952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_308.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_146_bfp.out9_18_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 583362752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_309.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_32_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 583362880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_310.const", + "file_size": 128 + }, + "onnx::MatMul_16920_11_0_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 583363008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_311.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 585231808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_312.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 585234368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_313.const", + "file_size": 2560 + }, + "onnx::MatMul_16921_11_0_31_12_6_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 585236928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_314.const", + "file_size": 7475200 + }, + "onnx::MatMul_16921_11_0_31_12_6_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 592712128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_315.const", + "file_size": 7475200 + }, + "onnx::MatMul_16922_11_0_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 600187328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_316.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 607662528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_317.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 607665088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_318.const", + "file_size": 2560 + }, + "Attention_7_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 607667648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_319.const", + "file_size": 1868800 + }, + "Attention_7_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 609536448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_320.const", + "file_size": 1868800 + }, + "Attention_7_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 611405248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_321.const", + "file_size": 1868800 + }, + "onnx::MatMul_16933_11_0_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 613274048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_322.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 615142848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_323.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 615145408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_324.const", + "file_size": 2560 + }, + "onnx::MatMul_16934": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 615147968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_325.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_147_bfp.out9_21_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 617016768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_326.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_36_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 617016896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_327.const", + "file_size": 128 + }, + "onnx::MatMul_16944_11_0_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 617017024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_328.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 618885824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_329.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 618888384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_330.const", + "file_size": 2560 + }, + "onnx::MatMul_16945_11_0_35_12_7_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 618890944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_331.const", + "file_size": 7475200 + }, + "onnx::MatMul_16945_11_0_35_12_7_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 626366144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_332.const", + "file_size": 7475200 + }, + "onnx::MatMul_16946_11_0_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 633841344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_333.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 641316544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_334.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 641319104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_335.const", + "file_size": 2560 + }, + "Attention_8_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 641321664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_336.const", + "file_size": 1868800 + }, + "Attention_8_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 643190464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_337.const", + "file_size": 1868800 + }, + "Attention_8_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 645059264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_338.const", + "file_size": 1868800 + }, + "onnx::MatMul_16957_11_0_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 646928064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_339.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 648796864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_340.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 648799424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_341.const", + "file_size": 2560 + }, + "onnx::MatMul_16958": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 648801984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_342.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_148_bfp.out9_24_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 650670784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_343.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_40_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 650670912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_344.const", + "file_size": 128 + }, + "onnx::MatMul_16968_11_0_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 650671040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_345.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 652539840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_346.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 652542400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_347.const", + "file_size": 2560 + }, + "onnx::MatMul_16969_11_0_39_12_8_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 652544960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_348.const", + "file_size": 7475200 + }, + "onnx::MatMul_16969_11_0_39_12_8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 660020160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_349.const", + "file_size": 7475200 + }, + "onnx::MatMul_16970_11_0_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 667495360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_350.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 674970560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_351.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 674973120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_352.const", + "file_size": 2560 + }, + "Attention_9_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 674975680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_353.const", + "file_size": 1868800 + }, + "Attention_9_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 676844480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_354.const", + "file_size": 1868800 + }, + "Attention_9_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 678713280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_355.const", + "file_size": 1868800 + }, + "onnx::MatMul_16981_11_0_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 680582080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_356.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 682450880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_357.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 682453440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_358.const", + "file_size": 2560 + }, + "onnx::MatMul_16982": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 682456000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_359.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_149_bfp.out9_27_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 684324800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_360.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_44_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 684324928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_361.const", + "file_size": 128 + }, + "onnx::MatMul_16992_11_0_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 684325056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_362.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 686193856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_363.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 686196416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_364.const", + "file_size": 2560 + }, + "onnx::MatMul_16993_11_0_43_12_9_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 686198976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_365.const", + "file_size": 7475200 + }, + "onnx::MatMul_16993_11_0_43_12_9_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 693674176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_366.const", + "file_size": 7475200 + }, + "onnx::MatMul_16994_11_0_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 701149376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_367.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 708624576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_368.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 708627136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_369.const", + "file_size": 2560 + }, + "Attention_10_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 708629696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_370.const", + "file_size": 1868800 + }, + "Attention_10_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 710498496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_371.const", + "file_size": 1868800 + }, + "Attention_10_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 712367296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_372.const", + "file_size": 1868800 + }, + "onnx::MatMul_17005_11_0_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 714236096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_373.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 716104896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_374.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 716107456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_375.const", + "file_size": 2560 + }, + "onnx::MatMul_17006": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 716110016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_376.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_150_bfp.out9_30_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 717978816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_377.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_48_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 717978944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_378.const", + "file_size": 128 + }, + "onnx::MatMul_17016_11_0_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 717979072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_379.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 719847872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_380.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 719850432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_381.const", + "file_size": 2560 + }, + "onnx::MatMul_17017_11_0_47_12_10_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 719852992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_382.const", + "file_size": 7475200 + }, + "onnx::MatMul_17017_11_0_47_12_10_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 727328192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_383.const", + "file_size": 7475200 + }, + "onnx::MatMul_17018_11_0_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 734803392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_384.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 742278592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_385.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 742281152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_386.const", + "file_size": 2560 + }, + "Attention_11_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 742283712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_387.const", + "file_size": 1868800 + }, + "Attention_11_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 744152512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_388.const", + "file_size": 1868800 + }, + "Attention_11_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 746021312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_389.const", + "file_size": 1868800 + }, + "onnx::MatMul_17029_11_0_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 747890112, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_390.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 749758912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_391.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 749761472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_392.const", + "file_size": 2560 + }, + "onnx::MatMul_17030": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 749764032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_393.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_151_bfp.out9_33_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 751632832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_394.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_52_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 751632960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_395.const", + "file_size": 128 + }, + "onnx::MatMul_17040_11_0_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 751633088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_396.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 753501888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_397.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 753504448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_398.const", + "file_size": 2560 + }, + "onnx::MatMul_17041_11_0_51_12_11_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 753507008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_399.const", + "file_size": 7475200 + }, + "onnx::MatMul_17041_11_0_51_12_11_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 760982208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_400.const", + "file_size": 7475200 + }, + "onnx::MatMul_17042_11_0_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 768457408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_401.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 775932608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_402.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 775935168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_403.const", + "file_size": 2560 + }, + "Attention_12_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 775937728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_404.const", + "file_size": 1868800 + }, + "Attention_12_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 777806528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_405.const", + "file_size": 1868800 + }, + "Attention_12_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 779675328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_406.const", + "file_size": 1868800 + }, + "onnx::MatMul_17053_11_0_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 781544128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_407.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 783412928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_408.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 783415488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_409.const", + "file_size": 2560 + }, + "onnx::MatMul_17054": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 783418048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_410.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_152_bfp.out9_36_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 785286848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_411.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_56_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 785286976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_412.const", + "file_size": 128 + }, + "onnx::MatMul_17064_11_0_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 785287104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_413.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 787155904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_414.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 787158464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_415.const", + "file_size": 2560 + }, + "onnx::MatMul_17065_11_0_55_12_12_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 787161024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_416.const", + "file_size": 7475200 + }, + "onnx::MatMul_17065_11_0_55_12_12_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 794636224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_417.const", + "file_size": 7475200 + }, + "onnx::MatMul_17066_11_0_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 802111424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_418.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 809586624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_419.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 809589184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_420.const", + "file_size": 2560 + }, + "Attention_13_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 809591744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_421.const", + "file_size": 1868800 + }, + "Attention_13_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 811460544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_422.const", + "file_size": 1868800 + }, + "Attention_13_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 813329344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_423.const", + "file_size": 1868800 + }, + "onnx::MatMul_17077_11_0_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 815198144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_424.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 817066944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_425.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 817069504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_426.const", + "file_size": 2560 + }, + "onnx::MatMul_17078": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 817072064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_427.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_153_bfp.out9_39_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 818940864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_428.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_60_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 818940992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_429.const", + "file_size": 128 + }, + "onnx::MatMul_17088_11_0_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 818941120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_430.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 820809920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_431.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.0.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 820812480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_432.const", + "file_size": 2560 + }, + "onnx::MatMul_17089_11_0_59_12_13_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 820815040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_433.const", + "file_size": 7475200 + }, + "onnx::MatMul_17089_11_0_59_12_13_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 828290240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_434.const", + "file_size": 7475200 + }, + "onnx::MatMul_17090_11_0_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 835765440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_435.const", + "file_size": 7475200 + }, + "onnx::MatMul_17091_11_0_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 843240640, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_436.const", + "file_size": 1868800 + }, + "GroupNorm_13_wts_6_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 845109440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_437.const", + "file_size": 5120 + }, + "NhwcConv_15_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 845114560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_438.const", + "file_size": 17244160 + }, + "GroupNorm_14_wts_6_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 862358720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_439.const", + "file_size": 5120 + }, + "NhwcConv_16_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 862363840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_440.const", + "file_size": 17244160 + }, + "GroupNorm_15_wts_6_1_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 879608000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_441.const", + "file_size": 5120 + }, + "onnx::MatMul_17098_11_0_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 879613120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_442.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 881481920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_443.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 881484480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_444.const", + "file_size": 2560 + }, + "Attention_14_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 881487040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_445.const", + "file_size": 1868800 + }, + "Attention_14_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 883355840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_446.const", + "file_size": 1868800 + }, + "Attention_14_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 885224640, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_447.const", + "file_size": 1868800 + }, + "onnx::MatMul_17109_11_0_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 887093440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_448.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 888962240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_449.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 888964800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_450.const", + "file_size": 2560 + }, + "onnx::MatMul_17110": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 888967360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_451.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_154_bfp.out9_42_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 890836160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_452.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_65_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 890836288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_453.const", + "file_size": 128 + }, + "onnx::MatMul_17120_11_0_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 890836416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_454.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 892705216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_455.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 892707776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_456.const", + "file_size": 2560 + }, + "onnx::MatMul_17121_11_0_65_12_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 892710336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_457.const", + "file_size": 7475200 + }, + "onnx::MatMul_17121_11_0_65_12_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 900185536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_458.const", + "file_size": 7475200 + }, + "onnx::MatMul_17122_11_0_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 907660736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_459.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 915135936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_460.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 915138496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_461.const", + "file_size": 2560 + }, + "Attention_15_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 915141056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_462.const", + "file_size": 1868800 + }, + "Attention_15_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 917009856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_463.const", + "file_size": 1868800 + }, + "Attention_15_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 918878656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_464.const", + "file_size": 1868800 + }, + "onnx::MatMul_17133_11_0_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 920747456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_465.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 922616256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_466.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 922618816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_467.const", + "file_size": 2560 + }, + "onnx::MatMul_17134": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 922621376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_468.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_155_bfp.out9_45_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 924490176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_469.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_69_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 924490304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_470.const", + "file_size": 128 + }, + "onnx::MatMul_17144_11_0_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 924490432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_471.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 926359232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_472.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 926361792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_473.const", + "file_size": 2560 + }, + "onnx::MatMul_17145_11_0_69_12_15_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 926364352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_474.const", + "file_size": 7475200 + }, + "onnx::MatMul_17145_11_0_69_12_15_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 933839552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_475.const", + "file_size": 7475200 + }, + "onnx::MatMul_17146_11_0_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 941314752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_476.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 948789952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_477.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 948792512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_478.const", + "file_size": 2560 + }, + "Attention_16_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 948795072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_479.const", + "file_size": 1868800 + }, + "Attention_16_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 950663872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_480.const", + "file_size": 1868800 + }, + "Attention_16_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 952532672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_481.const", + "file_size": 1868800 + }, + "onnx::MatMul_17157_11_0_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 954401472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_482.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 956270272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_483.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 956272832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_484.const", + "file_size": 2560 + }, + "onnx::MatMul_17158": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 956275392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_485.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_156_bfp.out9_48_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 958144192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_486.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_73_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 958144320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_487.const", + "file_size": 128 + }, + "onnx::MatMul_17168_11_0_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 958144448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_488.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 960013248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_489.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 960015808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_490.const", + "file_size": 2560 + }, + "onnx::MatMul_17169_11_0_73_12_16_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 960018368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_491.const", + "file_size": 7475200 + }, + "onnx::MatMul_17169_11_0_73_12_16_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 967493568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_492.const", + "file_size": 7475200 + }, + "onnx::MatMul_17170_11_0_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 974968768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_493.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 982443968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_494.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 982446528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_495.const", + "file_size": 2560 + }, + "Attention_17_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 982449088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_496.const", + "file_size": 1868800 + }, + "Attention_17_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 984317888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_497.const", + "file_size": 1868800 + }, + "Attention_17_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 986186688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_498.const", + "file_size": 1868800 + }, + "onnx::MatMul_17181_11_0_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 988055488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_499.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 989924288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_500.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 989926848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_501.const", + "file_size": 2560 + }, + "onnx::MatMul_17182": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 989929408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_502.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_157_bfp.out9_51_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 991798208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_503.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_77_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 991798336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_504.const", + "file_size": 128 + }, + "onnx::MatMul_17192_11_0_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 991798464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_505.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 993667264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_506.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 993669824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_507.const", + "file_size": 2560 + }, + "onnx::MatMul_17193_11_0_77_12_17_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 993672384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_508.const", + "file_size": 7475200 + }, + "onnx::MatMul_17193_11_0_77_12_17_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1001147584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_509.const", + "file_size": 7475200 + }, + "onnx::MatMul_17194_11_0_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1008622784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_510.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1016097984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_511.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1016100544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_512.const", + "file_size": 2560 + }, + "Attention_18_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1016103104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_513.const", + "file_size": 1868800 + }, + "Attention_18_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1017971904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_514.const", + "file_size": 1868800 + }, + "Attention_18_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1019840704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_515.const", + "file_size": 1868800 + }, + "onnx::MatMul_17205_11_0_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1021709504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_516.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1023578304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_517.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1023580864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_518.const", + "file_size": 2560 + }, + "onnx::MatMul_17206": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1023583424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_519.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_158_bfp.out9_54_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1025452224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_520.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_81_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1025452352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_521.const", + "file_size": 128 + }, + "onnx::MatMul_17216_11_0_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1025452480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_522.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1027321280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_523.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1027323840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_524.const", + "file_size": 2560 + }, + "onnx::MatMul_17217_11_0_81_12_18_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1027326400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_525.const", + "file_size": 7475200 + }, + "onnx::MatMul_17217_11_0_81_12_18_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1034801600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_526.const", + "file_size": 7475200 + }, + "onnx::MatMul_17218_11_0_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1042276800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_527.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1049752000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_528.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1049754560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_529.const", + "file_size": 2560 + }, + "Attention_19_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1049757120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_530.const", + "file_size": 1868800 + }, + "Attention_19_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1051625920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_531.const", + "file_size": 1868800 + }, + "Attention_19_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1053494720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_532.const", + "file_size": 1868800 + }, + "onnx::MatMul_17229_11_0_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1055363520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_533.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1057232320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_534.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1057234880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_535.const", + "file_size": 2560 + }, + "onnx::MatMul_17230": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1057237440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_536.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_159_bfp.out9_57_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1059106240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_537.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_85_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1059106368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_538.const", + "file_size": 128 + }, + "onnx::MatMul_17240_11_0_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1059106496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_539.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1060975296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_540.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1060977856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_541.const", + "file_size": 2560 + }, + "onnx::MatMul_17241_11_0_85_12_19_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1060980416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_542.const", + "file_size": 7475200 + }, + "onnx::MatMul_17241_11_0_85_12_19_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1068455616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_543.const", + "file_size": 7475200 + }, + "onnx::MatMul_17242_11_0_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1075930816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_544.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1083406016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_545.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1083408576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_546.const", + "file_size": 2560 + }, + "Attention_20_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1083411136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_547.const", + "file_size": 1868800 + }, + "Attention_20_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1085279936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_548.const", + "file_size": 1868800 + }, + "Attention_20_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1087148736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_549.const", + "file_size": 1868800 + }, + "onnx::MatMul_17253_11_0_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1089017536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_550.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1090886336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_551.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1090888896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_552.const", + "file_size": 2560 + }, + "onnx::MatMul_17254": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1090891456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_553.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_160_bfp.out9_60_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1092760256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_554.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_89_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1092760384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_555.const", + "file_size": 128 + }, + "onnx::MatMul_17264_11_0_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1092760512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_556.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1094629312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_557.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1094631872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_558.const", + "file_size": 2560 + }, + "onnx::MatMul_17265_11_0_89_12_20_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1094634432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_559.const", + "file_size": 7475200 + }, + "onnx::MatMul_17265_11_0_89_12_20_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1102109632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_560.const", + "file_size": 7475200 + }, + "onnx::MatMul_17266_11_0_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1109584832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_561.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1117060032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_562.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1117062592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_563.const", + "file_size": 2560 + }, + "Attention_21_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1117065152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_564.const", + "file_size": 1868800 + }, + "Attention_21_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1118933952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_565.const", + "file_size": 1868800 + }, + "Attention_21_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1120802752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_566.const", + "file_size": 1868800 + }, + "onnx::MatMul_17277_11_0_91": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1122671552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_567.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1124540352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_568.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1124542912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_569.const", + "file_size": 2560 + }, + "onnx::MatMul_17278": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1124545472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_570.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_161_bfp.out9_63_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1126414272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_571.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_93_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1126414400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_572.const", + "file_size": 128 + }, + "onnx::MatMul_17288_11_0_92": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1126414528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_573.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1128283328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_574.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1128285888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_575.const", + "file_size": 2560 + }, + "onnx::MatMul_17289_11_0_93_12_21_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1128288448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_576.const", + "file_size": 7475200 + }, + "onnx::MatMul_17289_11_0_93_12_21_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1135763648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_577.const", + "file_size": 7475200 + }, + "onnx::MatMul_17290_11_0_94": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1143238848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_578.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1150714048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_579.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1150716608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_580.const", + "file_size": 2560 + }, + "Attention_22_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1150719168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_581.const", + "file_size": 1868800 + }, + "Attention_22_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1152587968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_582.const", + "file_size": 1868800 + }, + "Attention_22_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1154456768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_583.const", + "file_size": 1868800 + }, + "onnx::MatMul_17301_11_0_95": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1156325568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_584.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1158194368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_585.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1158196928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_586.const", + "file_size": 2560 + }, + "onnx::MatMul_17302": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1158199488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_587.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_162_bfp.out9_66_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1160068288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_588.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_97_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1160068416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_589.const", + "file_size": 128 + }, + "onnx::MatMul_17312_11_0_96": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1160068544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_590.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1161937344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_591.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1161939904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_592.const", + "file_size": 2560 + }, + "onnx::MatMul_17313_11_0_97_12_22_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1161942464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_593.const", + "file_size": 7475200 + }, + "onnx::MatMul_17313_11_0_97_12_22_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1169417664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_594.const", + "file_size": 7475200 + }, + "onnx::MatMul_17314_11_0_98": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1176892864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_595.const", + "file_size": 7475200 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1184368064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_596.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1184370624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_597.const", + "file_size": 2560 + }, + "Attention_23_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1184373184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_598.const", + "file_size": 1868800 + }, + "Attention_23_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1186241984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_599.const", + "file_size": 1868800 + }, + "Attention_23_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1188110784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_600.const", + "file_size": 1868800 + }, + "onnx::MatMul_17325_11_0_99": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1189979584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_601.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1191848384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_602.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1191850944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_603.const", + "file_size": 2560 + }, + "onnx::MatMul_17326": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1191853504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_604.const", + "file_size": 1868800 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_163_bfp.out9_69_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1193722304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_605.const", + "file_size": 128 + }, + "/unet/down_blocks.2/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_101_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1193722432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_606.const", + "file_size": 128 + }, + "onnx::MatMul_17336_11_0_100": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1193722560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_607.const", + "file_size": 1868800 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1195591360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_608.const", + "file_size": 2560 + }, + "unet.down_blocks.2.attentions.1.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1195593920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_609.const", + "file_size": 2560 + }, + "onnx::MatMul_17337_11_0_101_12_23_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1195596480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_610.const", + "file_size": 7475200 + }, + "onnx::MatMul_17337_11_0_101_12_23_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1203071680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_611.const", + "file_size": 7475200 + }, + "onnx::MatMul_17338_11_0_102": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1210546880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_612.const", + "file_size": 7475200 + }, + "onnx::MatMul_17339_11_0_103": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1218022080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_613.const", + "file_size": 1868800 + }, + "GroupNorm_16_wts_6_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1219890880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_614.const", + "file_size": 5120 + }, + "NhwcConv_17_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 1219896000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_615.const", + "file_size": 17244160 + }, + "GroupNorm_17_wts_6_0_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1237140160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_616.const", + "file_size": 5120 + }, + "NhwcConv_18_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 1237145280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_617.const", + "file_size": 17244160 + }, + "GroupNorm_18_wts_6_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1254389440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_618.const", + "file_size": 5120 + }, + "onnx::MatMul_17346_11_0_104": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1254394560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_619.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1256263360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_620.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1256265920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_621.const", + "file_size": 2560 + }, + "Attention_24_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1256268480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_622.const", + "file_size": 1868800 + }, + "Attention_24_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1258137280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_623.const", + "file_size": 1868800 + }, + "Attention_24_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1260006080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_624.const", + "file_size": 1868800 + }, + "onnx::MatMul_17357_11_0_105": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1261874880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_625.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1263743680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_626.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1263746240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_627.const", + "file_size": 2560 + }, + "onnx::MatMul_17358": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1263748800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_628.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_164_bfp.out9_72_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1265617600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_629.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_106_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1265617728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_630.const", + "file_size": 128 + }, + "onnx::MatMul_17368_11_0_106": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1265617856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_631.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1267486656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_632.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1267489216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_633.const", + "file_size": 2560 + }, + "onnx::MatMul_17369_11_0_107_12_24_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1267491776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_634.const", + "file_size": 7475200 + }, + "onnx::MatMul_17369_11_0_107_12_24_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1274966976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_635.const", + "file_size": 7475200 + }, + "onnx::MatMul_17370_11_0_108": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1282442176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_636.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1289917376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_637.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1289919936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_638.const", + "file_size": 2560 + }, + "Attention_25_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1289922496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_639.const", + "file_size": 1868800 + }, + "Attention_25_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1291791296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_640.const", + "file_size": 1868800 + }, + "Attention_25_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1293660096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_641.const", + "file_size": 1868800 + }, + "onnx::MatMul_17381_11_0_109": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1295528896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_642.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1297397696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_643.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1297400256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_644.const", + "file_size": 2560 + }, + "onnx::MatMul_17382": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1297402816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_645.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_165_bfp.out9_75_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1299271616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_646.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_110_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1299271744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_647.const", + "file_size": 128 + }, + "onnx::MatMul_17392_11_0_110": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1299271872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_648.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1301140672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_649.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1301143232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_650.const", + "file_size": 2560 + }, + "onnx::MatMul_17393_11_0_111_12_25_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1301145792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_651.const", + "file_size": 7475200 + }, + "onnx::MatMul_17393_11_0_111_12_25_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1308620992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_652.const", + "file_size": 7475200 + }, + "onnx::MatMul_17394_11_0_112": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1316096192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_653.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1323571392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_654.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1323573952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_655.const", + "file_size": 2560 + }, + "Attention_26_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1323576512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_656.const", + "file_size": 1868800 + }, + "Attention_26_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1325445312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_657.const", + "file_size": 1868800 + }, + "Attention_26_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1327314112, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_658.const", + "file_size": 1868800 + }, + "onnx::MatMul_17405_11_0_113": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1329182912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_659.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1331051712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_660.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1331054272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_661.const", + "file_size": 2560 + }, + "onnx::MatMul_17406": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1331056832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_662.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_166_bfp.out9_78_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1332925632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_663.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_114_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1332925760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_664.const", + "file_size": 128 + }, + "onnx::MatMul_17416_11_0_114": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1332925888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_665.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1334794688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_666.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1334797248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_667.const", + "file_size": 2560 + }, + "onnx::MatMul_17417_11_0_115_12_26_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1334799808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_668.const", + "file_size": 7475200 + }, + "onnx::MatMul_17417_11_0_115_12_26_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1342275008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_669.const", + "file_size": 7475200 + }, + "onnx::MatMul_17418_11_0_116": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1349750208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_670.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1357225408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_671.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1357227968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_672.const", + "file_size": 2560 + }, + "Attention_27_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1357230528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_673.const", + "file_size": 1868800 + }, + "Attention_27_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1359099328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_674.const", + "file_size": 1868800 + }, + "Attention_27_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1360968128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_675.const", + "file_size": 1868800 + }, + "onnx::MatMul_17429_11_0_117": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1362836928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_676.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1364705728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_677.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1364708288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_678.const", + "file_size": 2560 + }, + "onnx::MatMul_17430": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1364710848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_679.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_167_bfp.out9_81_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1366579648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_680.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_118_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1366579776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_681.const", + "file_size": 128 + }, + "onnx::MatMul_17440_11_0_118": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1366579904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_682.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1368448704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_683.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1368451264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_684.const", + "file_size": 2560 + }, + "onnx::MatMul_17441_11_0_119_12_27_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1368453824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_685.const", + "file_size": 7475200 + }, + "onnx::MatMul_17441_11_0_119_12_27_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1375929024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_686.const", + "file_size": 7475200 + }, + "onnx::MatMul_17442_11_0_120": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1383404224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_687.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1390879424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_688.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1390881984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_689.const", + "file_size": 2560 + }, + "Attention_28_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1390884544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_690.const", + "file_size": 1868800 + }, + "Attention_28_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1392753344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_691.const", + "file_size": 1868800 + }, + "Attention_28_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1394622144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_692.const", + "file_size": 1868800 + }, + "onnx::MatMul_17453_11_0_121": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1396490944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_693.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1398359744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_694.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1398362304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_695.const", + "file_size": 2560 + }, + "onnx::MatMul_17454": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1398364864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_696.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_168_bfp.out9_84_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1400233664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_697.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_122_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1400233792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_698.const", + "file_size": 128 + }, + "onnx::MatMul_17464_11_0_122": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1400233920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_699.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1402102720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_700.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1402105280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_701.const", + "file_size": 2560 + }, + "onnx::MatMul_17465_11_0_123_12_28_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1402107840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_702.const", + "file_size": 7475200 + }, + "onnx::MatMul_17465_11_0_123_12_28_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1409583040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_703.const", + "file_size": 7475200 + }, + "onnx::MatMul_17466_11_0_124": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1417058240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_704.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1424533440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_705.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1424536000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_706.const", + "file_size": 2560 + }, + "Attention_29_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1424538560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_707.const", + "file_size": 1868800 + }, + "Attention_29_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1426407360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_708.const", + "file_size": 1868800 + }, + "Attention_29_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1428276160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_709.const", + "file_size": 1868800 + }, + "onnx::MatMul_17477_11_0_125": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1430144960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_710.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1432013760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_711.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1432016320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_712.const", + "file_size": 2560 + }, + "onnx::MatMul_17478": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1432018880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_713.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_169_bfp.out9_87_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1433887680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_714.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_126_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1433887808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_715.const", + "file_size": 128 + }, + "onnx::MatMul_17488_11_0_126": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1433887936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_716.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1435756736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_717.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1435759296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_718.const", + "file_size": 2560 + }, + "onnx::MatMul_17489_11_0_127_12_29_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1435761856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_719.const", + "file_size": 7475200 + }, + "onnx::MatMul_17489_11_0_127_12_29_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1443237056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_720.const", + "file_size": 7475200 + }, + "onnx::MatMul_17490_11_0_128": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1450712256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_721.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1458187456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_722.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1458190016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_723.const", + "file_size": 2560 + }, + "Attention_30_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1458192576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_724.const", + "file_size": 1868800 + }, + "Attention_30_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1460061376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_725.const", + "file_size": 1868800 + }, + "Attention_30_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1461930176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_726.const", + "file_size": 1868800 + }, + "onnx::MatMul_17501_11_0_129": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1463798976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_727.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1465667776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_728.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1465670336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_729.const", + "file_size": 2560 + }, + "onnx::MatMul_17502": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1465672896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_730.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_170_bfp.out9_90_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1467541696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_731.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_130_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1467541824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_732.const", + "file_size": 128 + }, + "onnx::MatMul_17512_11_0_130": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1467541952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_733.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1469410752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_734.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1469413312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_735.const", + "file_size": 2560 + }, + "onnx::MatMul_17513_11_0_131_12_30_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1469415872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_736.const", + "file_size": 7475200 + }, + "onnx::MatMul_17513_11_0_131_12_30_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1476891072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_737.const", + "file_size": 7475200 + }, + "onnx::MatMul_17514_11_0_132": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1484366272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_738.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1491841472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_739.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1491844032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_740.const", + "file_size": 2560 + }, + "Attention_31_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1491846592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_741.const", + "file_size": 1868800 + }, + "Attention_31_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1493715392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_742.const", + "file_size": 1868800 + }, + "Attention_31_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1495584192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_743.const", + "file_size": 1868800 + }, + "onnx::MatMul_17525_11_0_133": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1497452992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_744.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1499321792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_745.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1499324352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_746.const", + "file_size": 2560 + }, + "onnx::MatMul_17526": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1499326912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_747.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_171_bfp.out9_93_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1501195712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_748.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_134_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1501195840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_749.const", + "file_size": 128 + }, + "onnx::MatMul_17536_11_0_134": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1501195968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_750.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1503064768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_751.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1503067328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_752.const", + "file_size": 2560 + }, + "onnx::MatMul_17537_11_0_135_12_31_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1503069888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_753.const", + "file_size": 7475200 + }, + "onnx::MatMul_17537_11_0_135_12_31_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1510545088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_754.const", + "file_size": 7475200 + }, + "onnx::MatMul_17538_11_0_136": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1518020288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_755.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1525495488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_756.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1525498048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_757.const", + "file_size": 2560 + }, + "Attention_32_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1525500608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_758.const", + "file_size": 1868800 + }, + "Attention_32_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1527369408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_759.const", + "file_size": 1868800 + }, + "Attention_32_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1529238208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_760.const", + "file_size": 1868800 + }, + "onnx::MatMul_17549_11_0_137": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1531107008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_761.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1532975808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_762.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1532978368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_763.const", + "file_size": 2560 + }, + "onnx::MatMul_17550": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1532980928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_764.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_172_bfp.out9_96_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1534849728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_765.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_138_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1534849856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_766.const", + "file_size": 128 + }, + "onnx::MatMul_17560_11_0_138": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1534849984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_767.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1536718784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_768.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1536721344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_769.const", + "file_size": 2560 + }, + "onnx::MatMul_17561_11_0_139_12_32_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1536723904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_770.const", + "file_size": 7475200 + }, + "onnx::MatMul_17561_11_0_139_12_32_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1544199104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_771.const", + "file_size": 7475200 + }, + "onnx::MatMul_17562_11_0_140": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1551674304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_772.const", + "file_size": 7475200 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1559149504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_773.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1559152064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_774.const", + "file_size": 2560 + }, + "Attention_33_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1559154624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_775.const", + "file_size": 1868800 + }, + "Attention_33_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1561023424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_776.const", + "file_size": 1868800 + }, + "Attention_33_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1562892224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_777.const", + "file_size": 1868800 + }, + "onnx::MatMul_17573_11_0_141": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1564761024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_778.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1566629824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_779.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1566632384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_780.const", + "file_size": 2560 + }, + "onnx::MatMul_17574": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1566634944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_781.const", + "file_size": 1868800 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_173_bfp.out9_99_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1568503744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_782.const", + "file_size": 128 + }, + "/unet/mid_block/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_142_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1568503872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_783.const", + "file_size": 128 + }, + "onnx::MatMul_17584_11_0_142": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1568504000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_784.const", + "file_size": 1868800 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1570372800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_785.const", + "file_size": 2560 + }, + "unet.mid_block.attentions.0.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1570375360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_786.const", + "file_size": 2560 + }, + "onnx::MatMul_17585_11_0_143_12_33_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1570377920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_787.const", + "file_size": 7475200 + }, + "onnx::MatMul_17585_11_0_143_12_33_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1577853120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_788.const", + "file_size": 7475200 + }, + "onnx::MatMul_17586_11_0_144": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1585328320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_789.const", + "file_size": 7475200 + }, + "onnx::MatMul_17587_11_0_145": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1592803520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_790.const", + "file_size": 1868800 + }, + "GroupNorm_19_wts_6_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1594672320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_791.const", + "file_size": 5120 + }, + "NhwcConv_19_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 1594677440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_792.const", + "file_size": 17244160 + }, + "GroupNorm_20_wts_6_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1611921600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_793.const", + "file_size": 5120 + }, + "NhwcConv_20_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 1611926720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_794.const", + "file_size": 17244160 + }, + "GroupNorm_21_wts_6_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "op_tensor_size": 10240, + "offset": 1629170880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_795.const", + "file_size": 10240 + }, + "NhwcConv_21_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 33832960 + ], + "size_in_bytes": 33832960, + "op_tensor_size": 33832960, + "offset": 1629181120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_796.const", + "file_size": 33832960 + }, + "GroupNorm_22_wts_6_0_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1663014080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_797.const", + "file_size": 5120 + }, + "NhwcConv_22_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 1663019200, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_798.const", + "file_size": 17244160 + }, + "NhwcConv_23_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4014080 + ], + "size_in_bytes": 4014080, + "op_tensor_size": 4014080, + "offset": 1680263360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_799.const", + "file_size": 4014080 + }, + "GroupNorm_23_wts_6_1_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 1684277440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_800.const", + "file_size": 5120 + }, + "onnx::MatMul_17598_11_0_146": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1684282560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_801.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1686151360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_802.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1686153920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_803.const", + "file_size": 2560 + }, + "Attention_34_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1686156480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_804.const", + "file_size": 1868800 + }, + "Attention_34_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1688025280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_805.const", + "file_size": 1868800 + }, + "Attention_34_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1689894080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_806.const", + "file_size": 1868800 + }, + "onnx::MatMul_17609_11_0_147": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1691762880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_807.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1693631680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_808.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1693634240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_809.const", + "file_size": 2560 + }, + "onnx::MatMul_17610": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1693636800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_810.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_174_bfp.out9_102_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1695505600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_811.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_147_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1695505728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_812.const", + "file_size": 128 + }, + "onnx::MatMul_17620_11_0_148": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1695505856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_813.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1697374656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_814.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1697377216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_815.const", + "file_size": 2560 + }, + "onnx::MatMul_17621_11_0_149_12_34_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1697379776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_816.const", + "file_size": 7475200 + }, + "onnx::MatMul_17621_11_0_149_12_34_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1704854976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_817.const", + "file_size": 7475200 + }, + "onnx::MatMul_17622_11_0_150": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1712330176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_818.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1719805376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_819.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1719807936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_820.const", + "file_size": 2560 + }, + "Attention_35_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1719810496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_821.const", + "file_size": 1868800 + }, + "Attention_35_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1721679296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_822.const", + "file_size": 1868800 + }, + "Attention_35_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1723548096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_823.const", + "file_size": 1868800 + }, + "onnx::MatMul_17633_11_0_151": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1725416896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_824.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1727285696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_825.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1727288256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_826.const", + "file_size": 2560 + }, + "onnx::MatMul_17634": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1727290816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_827.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_175_bfp.out9_105_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1729159616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_828.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_151_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1729159744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_829.const", + "file_size": 128 + }, + "onnx::MatMul_17644_11_0_152": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1729159872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_830.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1731028672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_831.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1731031232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_832.const", + "file_size": 2560 + }, + "onnx::MatMul_17645_11_0_153_12_35_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1731033792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_833.const", + "file_size": 7475200 + }, + "onnx::MatMul_17645_11_0_153_12_35_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1738508992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_834.const", + "file_size": 7475200 + }, + "onnx::MatMul_17646_11_0_154": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1745984192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_835.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1753459392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_836.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1753461952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_837.const", + "file_size": 2560 + }, + "Attention_36_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1753464512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_838.const", + "file_size": 1868800 + }, + "Attention_36_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1755333312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_839.const", + "file_size": 1868800 + }, + "Attention_36_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1757202112, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_840.const", + "file_size": 1868800 + }, + "onnx::MatMul_17657_11_0_155": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1759070912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_841.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1760939712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_842.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1760942272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_843.const", + "file_size": 2560 + }, + "onnx::MatMul_17658": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1760944832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_844.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_176_bfp.out9_108_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1762813632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_845.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_155_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1762813760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_846.const", + "file_size": 128 + }, + "onnx::MatMul_17668_11_0_156": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1762813888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_847.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1764682688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_848.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1764685248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_849.const", + "file_size": 2560 + }, + "onnx::MatMul_17669_11_0_157_12_36_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1764687808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_850.const", + "file_size": 7475200 + }, + "onnx::MatMul_17669_11_0_157_12_36_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1772163008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_851.const", + "file_size": 7475200 + }, + "onnx::MatMul_17670_11_0_158": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1779638208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_852.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1787113408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_853.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1787115968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_854.const", + "file_size": 2560 + }, + "Attention_37_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1787118528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_855.const", + "file_size": 1868800 + }, + "Attention_37_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1788987328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_856.const", + "file_size": 1868800 + }, + "Attention_37_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1790856128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_857.const", + "file_size": 1868800 + }, + "onnx::MatMul_17681_11_0_159": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1792724928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_858.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1794593728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_859.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1794596288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_860.const", + "file_size": 2560 + }, + "onnx::MatMul_17682": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1794598848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_861.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_177_bfp.out9_111_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1796467648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_862.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_159_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1796467776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_863.const", + "file_size": 128 + }, + "onnx::MatMul_17692_11_0_160": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1796467904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_864.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1798336704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_865.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1798339264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_866.const", + "file_size": 2560 + }, + "onnx::MatMul_17693_11_0_161_12_37_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1798341824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_867.const", + "file_size": 7475200 + }, + "onnx::MatMul_17693_11_0_161_12_37_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1805817024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_868.const", + "file_size": 7475200 + }, + "onnx::MatMul_17694_11_0_162": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1813292224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_869.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1820767424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_870.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1820769984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_871.const", + "file_size": 2560 + }, + "Attention_38_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1820772544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_872.const", + "file_size": 1868800 + }, + "Attention_38_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1822641344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_873.const", + "file_size": 1868800 + }, + "Attention_38_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1824510144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_874.const", + "file_size": 1868800 + }, + "onnx::MatMul_17705_11_0_163": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1826378944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_875.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1828247744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_876.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1828250304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_877.const", + "file_size": 2560 + }, + "onnx::MatMul_17706": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1828252864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_878.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_178_bfp.out9_114_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1830121664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_879.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_163_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1830121792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_880.const", + "file_size": 128 + }, + "onnx::MatMul_17716_11_0_164": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1830121920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_881.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1831990720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_882.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1831993280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_883.const", + "file_size": 2560 + }, + "onnx::MatMul_17717_11_0_165_12_38_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1831995840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_884.const", + "file_size": 7475200 + }, + "onnx::MatMul_17717_11_0_165_12_38_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1839471040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_885.const", + "file_size": 7475200 + }, + "onnx::MatMul_17718_11_0_166": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1846946240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_886.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1854421440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_887.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1854424000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_888.const", + "file_size": 2560 + }, + "Attention_39_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1854426560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_889.const", + "file_size": 1868800 + }, + "Attention_39_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1856295360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_890.const", + "file_size": 1868800 + }, + "Attention_39_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1858164160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_891.const", + "file_size": 1868800 + }, + "onnx::MatMul_17729_11_0_167": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1860032960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_892.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1861901760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_893.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1861904320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_894.const", + "file_size": 2560 + }, + "onnx::MatMul_17730": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1861906880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_895.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_179_bfp.out9_117_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1863775680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_896.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_167_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1863775808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_897.const", + "file_size": 128 + }, + "onnx::MatMul_17740_11_0_168": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1863775936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_898.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1865644736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_899.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1865647296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_900.const", + "file_size": 2560 + }, + "onnx::MatMul_17741_11_0_169_12_39_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1865649856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_901.const", + "file_size": 7475200 + }, + "onnx::MatMul_17741_11_0_169_12_39_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1873125056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_902.const", + "file_size": 7475200 + }, + "onnx::MatMul_17742_11_0_170": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1880600256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_903.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1888075456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_904.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1888078016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_905.const", + "file_size": 2560 + }, + "Attention_40_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1888080576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_906.const", + "file_size": 1868800 + }, + "Attention_40_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1889949376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_907.const", + "file_size": 1868800 + }, + "Attention_40_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1891818176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_908.const", + "file_size": 1868800 + }, + "onnx::MatMul_17753_11_0_171": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1893686976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_909.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1895555776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_910.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1895558336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_911.const", + "file_size": 2560 + }, + "onnx::MatMul_17754": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1895560896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_912.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_180_bfp.out9_120_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1897429696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_913.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_171_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1897429824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_914.const", + "file_size": 128 + }, + "onnx::MatMul_17764_11_0_172": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1897429952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_915.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1899298752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_916.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1899301312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_917.const", + "file_size": 2560 + }, + "onnx::MatMul_17765_11_0_173_12_40_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1899303872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_918.const", + "file_size": 7475200 + }, + "onnx::MatMul_17765_11_0_173_12_40_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1906779072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_919.const", + "file_size": 7475200 + }, + "onnx::MatMul_17766_11_0_174": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1914254272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_920.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1921729472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_921.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1921732032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_922.const", + "file_size": 2560 + }, + "Attention_41_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1921734592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_923.const", + "file_size": 1868800 + }, + "Attention_41_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1923603392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_924.const", + "file_size": 1868800 + }, + "Attention_41_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1925472192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_925.const", + "file_size": 1868800 + }, + "onnx::MatMul_17777_11_0_175": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1927340992, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_926.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1929209792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_927.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1929212352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_928.const", + "file_size": 2560 + }, + "onnx::MatMul_17778": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1929214912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_929.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_181_bfp.out9_123_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1931083712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_930.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_175_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1931083840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_931.const", + "file_size": 128 + }, + "onnx::MatMul_17788_11_0_176": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1931083968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_932.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1932952768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_933.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1932955328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_934.const", + "file_size": 2560 + }, + "onnx::MatMul_17789_11_0_177_12_41_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1932957888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_935.const", + "file_size": 7475200 + }, + "onnx::MatMul_17789_11_0_177_12_41_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1940433088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_936.const", + "file_size": 7475200 + }, + "onnx::MatMul_17790_11_0_178": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1947908288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_937.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1955383488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_938.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1955386048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_939.const", + "file_size": 2560 + }, + "Attention_42_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1955388608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_940.const", + "file_size": 1868800 + }, + "Attention_42_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1957257408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_941.const", + "file_size": 1868800 + }, + "Attention_42_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1959126208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_942.const", + "file_size": 1868800 + }, + "onnx::MatMul_17801_11_0_179": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1960995008, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_943.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1962863808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_944.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1962866368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_945.const", + "file_size": 2560 + }, + "onnx::MatMul_17802": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1962868928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_946.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_182_bfp.out9_126_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1964737728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_947.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_179_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1964737856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_948.const", + "file_size": 128 + }, + "onnx::MatMul_17812_11_0_180": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1964737984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_949.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1966606784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_950.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1966609344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_951.const", + "file_size": 2560 + }, + "onnx::MatMul_17813_11_0_181_12_42_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1966611904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_952.const", + "file_size": 7475200 + }, + "onnx::MatMul_17813_11_0_181_12_42_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1974087104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_953.const", + "file_size": 7475200 + }, + "onnx::MatMul_17814_11_0_182": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 1981562304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_954.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1989037504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_955.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1989040064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_956.const", + "file_size": 2560 + }, + "Attention_43_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1989042624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_957.const", + "file_size": 1868800 + }, + "Attention_43_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1990911424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_958.const", + "file_size": 1868800 + }, + "Attention_43_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1992780224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_959.const", + "file_size": 1868800 + }, + "onnx::MatMul_17825_11_0_183": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1994649024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_960.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1996517824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_961.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 1996520384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_962.const", + "file_size": 2560 + }, + "onnx::MatMul_17826": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1996522944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_963.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_183_bfp.out9_129_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1998391744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_964.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.0/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_183_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 1998391872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_965.const", + "file_size": 128 + }, + "onnx::MatMul_17836_11_0_184": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 1998392000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_966.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2000260800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_967.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.0.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2000263360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_968.const", + "file_size": 2560 + }, + "onnx::MatMul_17837_11_0_185_12_43_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2000265920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_969.const", + "file_size": 7475200 + }, + "onnx::MatMul_17837_11_0_185_12_43_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2007741120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_970.const", + "file_size": 7475200 + }, + "onnx::MatMul_17838_11_0_186": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2015216320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_971.const", + "file_size": 7475200 + }, + "onnx::MatMul_17839_11_0_187": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2022691520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_972.const", + "file_size": 1868800 + }, + "GroupNorm_24_wts_6_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "op_tensor_size": 10240, + "offset": 2024560320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_973.const", + "file_size": 10240 + }, + "NhwcConv_24_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 33832960 + ], + "size_in_bytes": 33832960, + "op_tensor_size": 33832960, + "offset": 2024570560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_974.const", + "file_size": 33832960 + }, + "GroupNorm_25_wts_6_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 2058403520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_975.const", + "file_size": 5120 + }, + "NhwcConv_25_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 2058408640, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_976.const", + "file_size": 17244160 + }, + "NhwcConv_26_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4014080 + ], + "size_in_bytes": 4014080, + "op_tensor_size": 4014080, + "offset": 2075652800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_977.const", + "file_size": 4014080 + }, + "GroupNorm_26_wts_6_1_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 2079666880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_978.const", + "file_size": 5120 + }, + "onnx::MatMul_17846_11_0_188": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2079672000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_979.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2081540800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_980.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2081543360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_981.const", + "file_size": 2560 + }, + "Attention_44_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2081545920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_982.const", + "file_size": 1868800 + }, + "Attention_44_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2083414720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_983.const", + "file_size": 1868800 + }, + "Attention_44_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2085283520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_984.const", + "file_size": 1868800 + }, + "onnx::MatMul_17857_11_0_189": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2087152320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_985.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2089021120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_986.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2089023680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_987.const", + "file_size": 2560 + }, + "onnx::MatMul_17858": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2089026240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_988.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_184_bfp.out9_132_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2090895040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_989.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_188_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2090895168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_990.const", + "file_size": 128 + }, + "onnx::MatMul_17868_11_0_190": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2090895296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_991.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2092764096, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_992.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2092766656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_993.const", + "file_size": 2560 + }, + "onnx::MatMul_17869_11_0_191_12_44_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2092769216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_994.const", + "file_size": 7475200 + }, + "onnx::MatMul_17869_11_0_191_12_44_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2100244416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_995.const", + "file_size": 7475200 + }, + "onnx::MatMul_17870_11_0_192": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2107719616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_996.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2115194816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_997.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2115197376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_998.const", + "file_size": 2560 + }, + "Attention_45_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2115199936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_999.const", + "file_size": 1868800 + }, + "Attention_45_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2117068736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1000.const", + "file_size": 1868800 + }, + "Attention_45_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2118937536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1001.const", + "file_size": 1868800 + }, + "onnx::MatMul_17881_11_0_193": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2120806336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1002.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2122675136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1003.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2122677696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1004.const", + "file_size": 2560 + }, + "onnx::MatMul_17882": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2122680256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1005.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_185_bfp.out9_135_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2124549056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1006.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_192_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2124549184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1007.const", + "file_size": 128 + }, + "onnx::MatMul_17892_11_0_194": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2124549312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1008.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2126418112, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1009.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2126420672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1010.const", + "file_size": 2560 + }, + "onnx::MatMul_17893_11_0_195_12_45_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2126423232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1011.const", + "file_size": 7475200 + }, + "onnx::MatMul_17893_11_0_195_12_45_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2133898432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1012.const", + "file_size": 7475200 + }, + "onnx::MatMul_17894_11_0_196": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2141373632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1013.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2148848832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1014.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2148851392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1015.const", + "file_size": 2560 + }, + "Attention_46_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2148853952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1016.const", + "file_size": 1868800 + }, + "Attention_46_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2150722752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1017.const", + "file_size": 1868800 + }, + "Attention_46_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2152591552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1018.const", + "file_size": 1868800 + }, + "onnx::MatMul_17905_11_0_197": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2154460352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1019.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2156329152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1020.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2156331712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1021.const", + "file_size": 2560 + }, + "onnx::MatMul_17906": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2156334272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1022.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_186_bfp.out9_138_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2158203072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1023.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_196_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2158203200, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1024.const", + "file_size": 128 + }, + "onnx::MatMul_17916_11_0_198": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2158203328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1025.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2160072128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1026.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2160074688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1027.const", + "file_size": 2560 + }, + "onnx::MatMul_17917_11_0_199_12_46_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2160077248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1028.const", + "file_size": 7475200 + }, + "onnx::MatMul_17917_11_0_199_12_46_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2167552448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1029.const", + "file_size": 7475200 + }, + "onnx::MatMul_17918_11_0_200": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2175027648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1030.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2182502848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1031.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2182505408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1032.const", + "file_size": 2560 + }, + "Attention_47_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2182507968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1033.const", + "file_size": 1868800 + }, + "Attention_47_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2184376768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1034.const", + "file_size": 1868800 + }, + "Attention_47_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2186245568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1035.const", + "file_size": 1868800 + }, + "onnx::MatMul_17929_11_0_201": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2188114368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1036.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2189983168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1037.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2189985728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1038.const", + "file_size": 2560 + }, + "onnx::MatMul_17930": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2189988288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1039.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_187_bfp.out9_141_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2191857088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1040.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_200_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2191857216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1041.const", + "file_size": 128 + }, + "onnx::MatMul_17940_11_0_202": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2191857344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1042.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2193726144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1043.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2193728704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1044.const", + "file_size": 2560 + }, + "onnx::MatMul_17941_11_0_203_12_47_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2193731264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1045.const", + "file_size": 7475200 + }, + "onnx::MatMul_17941_11_0_203_12_47_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2201206464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1046.const", + "file_size": 7475200 + }, + "onnx::MatMul_17942_11_0_204": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2208681664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1047.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2216156864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1048.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2216159424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1049.const", + "file_size": 2560 + }, + "Attention_48_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2216161984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1050.const", + "file_size": 1868800 + }, + "Attention_48_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2218030784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1051.const", + "file_size": 1868800 + }, + "Attention_48_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2219899584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1052.const", + "file_size": 1868800 + }, + "onnx::MatMul_17953_11_0_205": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2221768384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1053.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2223637184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1054.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2223639744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1055.const", + "file_size": 2560 + }, + "onnx::MatMul_17954": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2223642304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1056.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_188_bfp.out9_144_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2225511104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1057.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_204_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2225511232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1058.const", + "file_size": 128 + }, + "onnx::MatMul_17964_11_0_206": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2225511360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1059.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2227380160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1060.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2227382720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1061.const", + "file_size": 2560 + }, + "onnx::MatMul_17965_11_0_207_12_48_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2227385280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1062.const", + "file_size": 7475200 + }, + "onnx::MatMul_17965_11_0_207_12_48_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2234860480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1063.const", + "file_size": 7475200 + }, + "onnx::MatMul_17966_11_0_208": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2242335680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1064.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2249810880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1065.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2249813440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1066.const", + "file_size": 2560 + }, + "Attention_49_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2249816000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1067.const", + "file_size": 1868800 + }, + "Attention_49_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2251684800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1068.const", + "file_size": 1868800 + }, + "Attention_49_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2253553600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1069.const", + "file_size": 1868800 + }, + "onnx::MatMul_17977_11_0_209": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2255422400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1070.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2257291200, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1071.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2257293760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1072.const", + "file_size": 2560 + }, + "onnx::MatMul_17978": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2257296320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1073.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_189_bfp.out9_147_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2259165120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1074.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_208_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2259165248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1075.const", + "file_size": 128 + }, + "onnx::MatMul_17988_11_0_210": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2259165376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1076.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2261034176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1077.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2261036736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1078.const", + "file_size": 2560 + }, + "onnx::MatMul_17989_11_0_211_12_49_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2261039296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1079.const", + "file_size": 7475200 + }, + "onnx::MatMul_17989_11_0_211_12_49_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2268514496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1080.const", + "file_size": 7475200 + }, + "onnx::MatMul_17990_11_0_212": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2275989696, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1081.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2283464896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1082.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2283467456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1083.const", + "file_size": 2560 + }, + "Attention_50_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2283470016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1084.const", + "file_size": 1868800 + }, + "Attention_50_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2285338816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1085.const", + "file_size": 1868800 + }, + "Attention_50_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2287207616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1086.const", + "file_size": 1868800 + }, + "onnx::MatMul_18001_11_0_213": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2289076416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1087.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2290945216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1088.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2290947776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1089.const", + "file_size": 2560 + }, + "onnx::MatMul_18002": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2290950336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1090.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_190_bfp.out9_150_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2292819136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1091.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_212_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2292819264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1092.const", + "file_size": 128 + }, + "onnx::MatMul_18012_11_0_214": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2292819392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1093.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2294688192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1094.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2294690752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1095.const", + "file_size": 2560 + }, + "onnx::MatMul_18013_11_0_215_12_50_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2294693312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1096.const", + "file_size": 7475200 + }, + "onnx::MatMul_18013_11_0_215_12_50_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2302168512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1097.const", + "file_size": 7475200 + }, + "onnx::MatMul_18014_11_0_216": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2309643712, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1098.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2317118912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1099.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2317121472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1100.const", + "file_size": 2560 + }, + "Attention_51_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2317124032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1101.const", + "file_size": 1868800 + }, + "Attention_51_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2318992832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1102.const", + "file_size": 1868800 + }, + "Attention_51_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2320861632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1103.const", + "file_size": 1868800 + }, + "onnx::MatMul_18025_11_0_217": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2322730432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1104.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2324599232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1105.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2324601792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1106.const", + "file_size": 2560 + }, + "onnx::MatMul_18026": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2324604352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1107.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_191_bfp.out9_153_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2326473152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1108.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_216_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2326473280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1109.const", + "file_size": 128 + }, + "onnx::MatMul_18036_11_0_218": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2326473408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1110.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2328342208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1111.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2328344768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1112.const", + "file_size": 2560 + }, + "onnx::MatMul_18037_11_0_219_12_51_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2328347328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1113.const", + "file_size": 7475200 + }, + "onnx::MatMul_18037_11_0_219_12_51_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2335822528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1114.const", + "file_size": 7475200 + }, + "onnx::MatMul_18038_11_0_220": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2343297728, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1115.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2350772928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1116.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2350775488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1117.const", + "file_size": 2560 + }, + "Attention_52_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2350778048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1118.const", + "file_size": 1868800 + }, + "Attention_52_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2352646848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1119.const", + "file_size": 1868800 + }, + "Attention_52_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2354515648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1120.const", + "file_size": 1868800 + }, + "onnx::MatMul_18049_11_0_221": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2356384448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1121.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2358253248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1122.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2358255808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1123.const", + "file_size": 2560 + }, + "onnx::MatMul_18050": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2358258368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1124.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_192_bfp.out9_156_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2360127168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1125.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_220_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2360127296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1126.const", + "file_size": 128 + }, + "onnx::MatMul_18060_11_0_222": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2360127424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1127.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2361996224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1128.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2361998784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1129.const", + "file_size": 2560 + }, + "onnx::MatMul_18061_11_0_223_12_52_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2362001344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1130.const", + "file_size": 7475200 + }, + "onnx::MatMul_18061_11_0_223_12_52_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2369476544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1131.const", + "file_size": 7475200 + }, + "onnx::MatMul_18062_11_0_224": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2376951744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1132.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2384426944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1133.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2384429504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1134.const", + "file_size": 2560 + }, + "Attention_53_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2384432064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1135.const", + "file_size": 1868800 + }, + "Attention_53_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2386300864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1136.const", + "file_size": 1868800 + }, + "Attention_53_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2388169664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1137.const", + "file_size": 1868800 + }, + "onnx::MatMul_18073_11_0_225": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2390038464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1138.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2391907264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1139.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2391909824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1140.const", + "file_size": 2560 + }, + "onnx::MatMul_18074": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2391912384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1141.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_193_bfp.out9_159_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2393781184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1142.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.1/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_224_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2393781312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1143.const", + "file_size": 128 + }, + "onnx::MatMul_18084_11_0_226": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2393781440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1144.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2395650240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1145.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.1.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2395652800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1146.const", + "file_size": 2560 + }, + "onnx::MatMul_18085_11_0_227_12_53_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2395655360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1147.const", + "file_size": 7475200 + }, + "onnx::MatMul_18085_11_0_227_12_53_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2403130560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1148.const", + "file_size": 7475200 + }, + "onnx::MatMul_18086_11_0_228": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2410605760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1149.const", + "file_size": 7475200 + }, + "onnx::MatMul_18087_11_0_229": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2418080960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1150.const", + "file_size": 1868800 + }, + "GroupNorm_27_wts_6_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 3840 + ], + "size_in_bytes": 7680, + "op_tensor_size": 7680, + "offset": 2419949760, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1151.const", + "file_size": 7680 + }, + "NhwcConv_27_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 25374720 + ], + "size_in_bytes": 25374720, + "op_tensor_size": 25374720, + "offset": 2419957440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1152.const", + "file_size": 25374720 + }, + "GroupNorm_28_wts_6_0_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 2445332160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1153.const", + "file_size": 5120 + }, + "NhwcConv_28_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 2445337280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1154.const", + "file_size": 17244160 + }, + "NhwcConv_29_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3010560 + ], + "size_in_bytes": 3010560, + "op_tensor_size": 3010560, + "offset": 2462581440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1155.const", + "file_size": 3010560 + }, + "GroupNorm_29_wts_6_1_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 2465592000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1156.const", + "file_size": 5120 + }, + "onnx::MatMul_18094_11_0_230": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2465597120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1157.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2467465920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1158.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2467468480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1159.const", + "file_size": 2560 + }, + "Attention_54_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2467471040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1160.const", + "file_size": 1868800 + }, + "Attention_54_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2469339840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1161.const", + "file_size": 1868800 + }, + "Attention_54_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2471208640, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1162.const", + "file_size": 1868800 + }, + "onnx::MatMul_18105_11_0_231": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2473077440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1163.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2474946240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1164.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2474948800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1165.const", + "file_size": 2560 + }, + "onnx::MatMul_18106": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2474951360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1166.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_194_bfp.out9_162_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2476820160, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1167.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_229_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2476820288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1168.const", + "file_size": 128 + }, + "onnx::MatMul_18116_11_0_232": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2476820416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1169.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2478689216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1170.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2478691776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1171.const", + "file_size": 2560 + }, + "onnx::MatMul_18117_11_0_233_12_54_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2478694336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1172.const", + "file_size": 7475200 + }, + "onnx::MatMul_18117_11_0_233_12_54_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2486169536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1173.const", + "file_size": 7475200 + }, + "onnx::MatMul_18118_11_0_234": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2493644736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1174.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2501119936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1175.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2501122496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1176.const", + "file_size": 2560 + }, + "Attention_55_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2501125056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1177.const", + "file_size": 1868800 + }, + "Attention_55_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2502993856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1178.const", + "file_size": 1868800 + }, + "Attention_55_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2504862656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1179.const", + "file_size": 1868800 + }, + "onnx::MatMul_18129_11_0_235": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2506731456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1180.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2508600256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1181.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2508602816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1182.const", + "file_size": 2560 + }, + "onnx::MatMul_18130": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2508605376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1183.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_195_bfp.out9_165_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2510474176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1184.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_233_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2510474304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1185.const", + "file_size": 128 + }, + "onnx::MatMul_18140_11_0_236": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2510474432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1186.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2512343232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1187.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2512345792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1188.const", + "file_size": 2560 + }, + "onnx::MatMul_18141_11_0_237_12_55_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2512348352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1189.const", + "file_size": 7475200 + }, + "onnx::MatMul_18141_11_0_237_12_55_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2519823552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1190.const", + "file_size": 7475200 + }, + "onnx::MatMul_18142_11_0_238": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2527298752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1191.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2534773952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1192.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2534776512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1193.const", + "file_size": 2560 + }, + "Attention_56_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2534779072, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1194.const", + "file_size": 1868800 + }, + "Attention_56_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2536647872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1195.const", + "file_size": 1868800 + }, + "Attention_56_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2538516672, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1196.const", + "file_size": 1868800 + }, + "onnx::MatMul_18153_11_0_239": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2540385472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1197.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2542254272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1198.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2542256832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1199.const", + "file_size": 2560 + }, + "onnx::MatMul_18154": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2542259392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1200.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/to_q/MatMul_output_0.out2_6_196_bfp.out9_168_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2544128192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1201.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.2/attn2/Reshape_3_output_0.out2_3_237_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2544128320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1202.const", + "file_size": 128 + }, + "onnx::MatMul_18164_11_0_240": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2544128448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1203.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2545997248, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1204.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.2.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2545999808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1205.const", + "file_size": 2560 + }, + "onnx::MatMul_18165_11_0_241_12_56_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2546002368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1206.const", + "file_size": 7475200 + }, + "onnx::MatMul_18165_11_0_241_12_56_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2553477568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1207.const", + "file_size": 7475200 + }, + "onnx::MatMul_18166_11_0_242": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2560952768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1208.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2568427968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1209.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2568430528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1210.const", + "file_size": 2560 + }, + "Attention_57_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2568433088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1211.const", + "file_size": 1868800 + }, + "Attention_57_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2570301888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1212.const", + "file_size": 1868800 + }, + "Attention_57_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2572170688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1213.const", + "file_size": 1868800 + }, + "onnx::MatMul_18177_11_0_243": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2574039488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1214.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2575908288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1215.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2575910848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1216.const", + "file_size": 2560 + }, + "onnx::MatMul_18178": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2575913408, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1217.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/to_q/MatMul_output_0.out2_6_197_bfp.out9_171_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2577782208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1218.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.3/attn2/Reshape_3_output_0.out2_3_241_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2577782336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1219.const", + "file_size": 128 + }, + "onnx::MatMul_18188_11_0_244": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2577782464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1220.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2579651264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1221.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.3.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2579653824, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1222.const", + "file_size": 2560 + }, + "onnx::MatMul_18189_11_0_245_12_57_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2579656384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1223.const", + "file_size": 7475200 + }, + "onnx::MatMul_18189_11_0_245_12_57_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2587131584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1224.const", + "file_size": 7475200 + }, + "onnx::MatMul_18190_11_0_246": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2594606784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1225.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2602081984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1226.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2602084544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1227.const", + "file_size": 2560 + }, + "Attention_58_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2602087104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1228.const", + "file_size": 1868800 + }, + "Attention_58_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2603955904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1229.const", + "file_size": 1868800 + }, + "Attention_58_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2605824704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1230.const", + "file_size": 1868800 + }, + "onnx::MatMul_18201_11_0_247": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2607693504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1231.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2609562304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1232.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2609564864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1233.const", + "file_size": 2560 + }, + "onnx::MatMul_18202": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2609567424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1234.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/to_q/MatMul_output_0.out2_6_198_bfp.out9_174_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2611436224, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1235.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.4/attn2/Reshape_3_output_0.out2_3_245_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2611436352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1236.const", + "file_size": 128 + }, + "onnx::MatMul_18212_11_0_248": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2611436480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1237.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2613305280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1238.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.4.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2613307840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1239.const", + "file_size": 2560 + }, + "onnx::MatMul_18213_11_0_249_12_58_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2613310400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1240.const", + "file_size": 7475200 + }, + "onnx::MatMul_18213_11_0_249_12_58_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2620785600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1241.const", + "file_size": 7475200 + }, + "onnx::MatMul_18214_11_0_250": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2628260800, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1242.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2635736000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1243.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2635738560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1244.const", + "file_size": 2560 + }, + "Attention_59_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2635741120, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1245.const", + "file_size": 1868800 + }, + "Attention_59_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2637609920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1246.const", + "file_size": 1868800 + }, + "Attention_59_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2639478720, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1247.const", + "file_size": 1868800 + }, + "onnx::MatMul_18225_11_0_251": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2641347520, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1248.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2643216320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1249.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2643218880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1250.const", + "file_size": 2560 + }, + "onnx::MatMul_18226": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2643221440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1251.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/to_q/MatMul_output_0.out2_6_199_bfp.out9_177_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2645090240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1252.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.5/attn2/Reshape_3_output_0.out2_3_249_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2645090368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1253.const", + "file_size": 128 + }, + "onnx::MatMul_18236_11_0_252": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2645090496, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1254.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2646959296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1255.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.5.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2646961856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1256.const", + "file_size": 2560 + }, + "onnx::MatMul_18237_11_0_253_12_59_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2646964416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1257.const", + "file_size": 7475200 + }, + "onnx::MatMul_18237_11_0_253_12_59_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2654439616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1258.const", + "file_size": 7475200 + }, + "onnx::MatMul_18238_11_0_254": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2661914816, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1259.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2669390016, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1260.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2669392576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1261.const", + "file_size": 2560 + }, + "Attention_60_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2669395136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1262.const", + "file_size": 1868800 + }, + "Attention_60_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2671263936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1263.const", + "file_size": 1868800 + }, + "Attention_60_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2673132736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1264.const", + "file_size": 1868800 + }, + "onnx::MatMul_18249_11_0_255": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2675001536, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1265.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2676870336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1266.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2676872896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1267.const", + "file_size": 2560 + }, + "onnx::MatMul_18250": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2676875456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1268.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/to_q/MatMul_output_0.out2_6_200_bfp.out9_180_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2678744256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1269.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.6/attn2/Reshape_3_output_0.out2_3_253_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2678744384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1270.const", + "file_size": 128 + }, + "onnx::MatMul_18260_11_0_256": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2678744512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1271.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2680613312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1272.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.6.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2680615872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1273.const", + "file_size": 2560 + }, + "onnx::MatMul_18261_11_0_257_12_60_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2680618432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1274.const", + "file_size": 7475200 + }, + "onnx::MatMul_18261_11_0_257_12_60_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2688093632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1275.const", + "file_size": 7475200 + }, + "onnx::MatMul_18262_11_0_258": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2695568832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1276.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2703044032, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1277.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2703046592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1278.const", + "file_size": 2560 + }, + "Attention_61_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2703049152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1279.const", + "file_size": 1868800 + }, + "Attention_61_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2704917952, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1280.const", + "file_size": 1868800 + }, + "Attention_61_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2706786752, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1281.const", + "file_size": 1868800 + }, + "onnx::MatMul_18273_11_0_259": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2708655552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1282.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2710524352, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1283.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2710526912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1284.const", + "file_size": 2560 + }, + "onnx::MatMul_18274": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2710529472, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1285.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/to_q/MatMul_output_0.out2_6_201_bfp.out9_183_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2712398272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1286.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.7/attn2/Reshape_3_output_0.out2_3_257_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2712398400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1287.const", + "file_size": 128 + }, + "onnx::MatMul_18284_11_0_260": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2712398528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1288.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2714267328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1289.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.7.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2714269888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1290.const", + "file_size": 2560 + }, + "onnx::MatMul_18285_11_0_261_12_61_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2714272448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1291.const", + "file_size": 7475200 + }, + "onnx::MatMul_18285_11_0_261_12_61_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2721747648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1292.const", + "file_size": 7475200 + }, + "onnx::MatMul_18286_11_0_262": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2729222848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1293.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2736698048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1294.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2736700608, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1295.const", + "file_size": 2560 + }, + "Attention_62_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2736703168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1296.const", + "file_size": 1868800 + }, + "Attention_62_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2738571968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1297.const", + "file_size": 1868800 + }, + "Attention_62_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2740440768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1298.const", + "file_size": 1868800 + }, + "onnx::MatMul_18297_11_0_263": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2742309568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1299.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2744178368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1300.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2744180928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1301.const", + "file_size": 2560 + }, + "onnx::MatMul_18298": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2744183488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1302.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/to_q/MatMul_output_0.out2_6_202_bfp.out9_186_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2746052288, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1303.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.8/attn2/Reshape_3_output_0.out2_3_261_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2746052416, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1304.const", + "file_size": 128 + }, + "onnx::MatMul_18308_11_0_264": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2746052544, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1305.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2747921344, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1306.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.8.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2747923904, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1307.const", + "file_size": 2560 + }, + "onnx::MatMul_18309_11_0_265_12_62_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2747926464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1308.const", + "file_size": 7475200 + }, + "onnx::MatMul_18309_11_0_265_12_62_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2755401664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1309.const", + "file_size": 7475200 + }, + "onnx::MatMul_18310_11_0_266": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2762876864, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1310.const", + "file_size": 7475200 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2770352064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1311.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2770354624, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1312.const", + "file_size": 2560 + }, + "Attention_63_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2770357184, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1313.const", + "file_size": 1868800 + }, + "Attention_63_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2772225984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1314.const", + "file_size": 1868800 + }, + "Attention_63_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2774094784, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1315.const", + "file_size": 1868800 + }, + "onnx::MatMul_18321_11_0_267": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2775963584, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1316.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2777832384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1317.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2777834944, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1318.const", + "file_size": 2560 + }, + "onnx::MatMul_18322": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2777837504, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1319.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/to_q/MatMul_output_0.out2_6_203_bfp.out9_189_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2779706304, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1320.const", + "file_size": 128 + }, + "/unet/up_blocks.0/attentions.2/transformer_blocks.9/attn2/Reshape_3_output_0.out2_3_265_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2779706432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1321.const", + "file_size": 128 + }, + "onnx::MatMul_18332_11_0_268": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2779706560, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1322.const", + "file_size": 1868800 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2781575360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1323.const", + "file_size": 2560 + }, + "unet.up_blocks.0.attentions.2.transformer_blocks.9.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2781577920, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1324.const", + "file_size": 2560 + }, + "onnx::MatMul_18333_11_0_269_12_63_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2781580480, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1325.const", + "file_size": 7475200 + }, + "onnx::MatMul_18333_11_0_269_12_63_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2789055680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1326.const", + "file_size": 7475200 + }, + "onnx::MatMul_18334_11_0_270": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "op_tensor_size": 7475200, + "offset": 2796530880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1327.const", + "file_size": 7475200 + }, + "onnx::MatMul_18335_11_0_271": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2804006080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1328.const", + "file_size": 1868800 + }, + "/unet/up_blocks.0/upsamplers.0/Resize.weights3_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 2805874880, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1329.const", + "file_size": 256 + }, + "NhwcConv_30_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "op_tensor_size": 17244160, + "offset": 2805875136, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1330.const", + "file_size": 17244160 + }, + "GroupNorm_30_wts_6_0_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 3840 + ], + "size_in_bytes": 7680, + "op_tensor_size": 7680, + "offset": 2823119296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1331.const", + "file_size": 7680 + }, + "NhwcConv_31_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 12851200 + ], + "size_in_bytes": 12851200, + "op_tensor_size": 12851200, + "offset": 2823126976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1332.const", + "file_size": 12851200 + }, + "GroupNorm_31_wts_6_0_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2835978176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1333.const", + "file_size": 2560 + }, + "NhwcConv_32_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 2835980736, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1334.const", + "file_size": 4311040 + }, + "NhwcConv_33_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1505280 + ], + "size_in_bytes": 1505280, + "op_tensor_size": 1505280, + "offset": 2840291776, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1335.const", + "file_size": 1505280 + }, + "GroupNorm_32_wts_6_1_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2841797056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1336.const", + "file_size": 2560 + }, + "onnx::MatMul_18343_11_0_272": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2841799616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1337.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2842270656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1338.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2842271936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1339.const", + "file_size": 1280 + }, + "Attention_64_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2842273216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1340.const", + "file_size": 471040 + }, + "Attention_64_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2842744256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1341.const", + "file_size": 471040 + }, + "Attention_64_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2843215296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1342.const", + "file_size": 471040 + }, + "onnx::MatMul_18354_11_0_273": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2843686336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1343.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2844157376, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1344.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2844158656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1345.const", + "file_size": 1280 + }, + "onnx::MatMul_18355": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2844159936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1346.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_204_bfp.out9_192_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2844630976, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1347.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_270_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2844631104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1348.const", + "file_size": 128 + }, + "onnx::MatMul_18365_11_0_274": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2844631232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1349.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2845102272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1350.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2845103552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1351.const", + "file_size": 1280 + }, + "onnx::MatMul_18366_11_0_275_12_64_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2845104832, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1352.const", + "file_size": 1868800 + }, + "onnx::MatMul_18366_11_0_275_12_64_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2846973632, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1353.const", + "file_size": 1868800 + }, + "onnx::MatMul_18367_11_0_276": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2848842432, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1354.const", + "file_size": 1884160 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2850726592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1355.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2850727872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1356.const", + "file_size": 1280 + }, + "Attention_65_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2850729152, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1357.const", + "file_size": 471040 + }, + "Attention_65_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2851200192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1358.const", + "file_size": 471040 + }, + "Attention_65_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2851671232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1359.const", + "file_size": 471040 + }, + "onnx::MatMul_18378_11_0_277": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2852142272, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1360.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2852613312, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1361.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2852614592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1362.const", + "file_size": 1280 + }, + "onnx::MatMul_18379": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2852615872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1363.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_205_bfp.out9_195_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2853086912, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1364.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.0/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_274_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2853087040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1365.const", + "file_size": 128 + }, + "onnx::MatMul_18389_11_0_278": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2853087168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1366.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2853558208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1367.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.0.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2853559488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1368.const", + "file_size": 1280 + }, + "onnx::MatMul_18390_11_0_279_12_65_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2853560768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1369.const", + "file_size": 1868800 + }, + "onnx::MatMul_18390_11_0_279_12_65_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2855429568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1370.const", + "file_size": 1868800 + }, + "onnx::MatMul_18391_11_0_280": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2857298368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1371.const", + "file_size": 1884160 + }, + "onnx::MatMul_18392_11_0_281": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2859182528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1372.const", + "file_size": 471040 + }, + "GroupNorm_33_wts_6_0_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "op_tensor_size": 5120, + "offset": 2859653568, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1373.const", + "file_size": 5120 + }, + "NhwcConv_34_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 8622080 + ], + "size_in_bytes": 8622080, + "op_tensor_size": 8622080, + "offset": 2859658688, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1374.const", + "file_size": 8622080 + }, + "GroupNorm_34_wts_6_0_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2868280768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1375.const", + "file_size": 2560 + }, + "NhwcConv_35_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 2868283328, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1376.const", + "file_size": 4311040 + }, + "NhwcConv_36_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1003520 + ], + "size_in_bytes": 1003520, + "op_tensor_size": 1003520, + "offset": 2872594368, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1377.const", + "file_size": 1003520 + }, + "GroupNorm_35_wts_6_1_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2873597888, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1378.const", + "file_size": 2560 + }, + "onnx::MatMul_18399_11_0_282": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2873600448, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1379.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2874071488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1380.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2874072768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1381.const", + "file_size": 1280 + }, + "Attention_66_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2874074048, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1382.const", + "file_size": 471040 + }, + "Attention_66_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2874545088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1383.const", + "file_size": 471040 + }, + "Attention_66_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2875016128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1384.const", + "file_size": 471040 + }, + "onnx::MatMul_18410_11_0_283": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2875487168, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1385.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2875958208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1386.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2875959488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1387.const", + "file_size": 1280 + }, + "onnx::MatMul_18411": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2875960768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1388.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_206_bfp.out9_198_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2876431808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1389.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_279_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2876431936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1390.const", + "file_size": 128 + }, + "onnx::MatMul_18421_11_0_284": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2876432064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1391.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2876903104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1392.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2876904384, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1393.const", + "file_size": 1280 + }, + "onnx::MatMul_18422_11_0_285_12_66_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2876905664, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1394.const", + "file_size": 1868800 + }, + "onnx::MatMul_18422_11_0_285_12_66_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2878774464, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1395.const", + "file_size": 1868800 + }, + "onnx::MatMul_18423_11_0_286": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2880643264, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1396.const", + "file_size": 1884160 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2882527424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1397.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2882528704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1398.const", + "file_size": 1280 + }, + "Attention_67_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2882529984, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1399.const", + "file_size": 471040 + }, + "Attention_67_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2883001024, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1400.const", + "file_size": 471040 + }, + "Attention_67_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2883472064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1401.const", + "file_size": 471040 + }, + "onnx::MatMul_18434_11_0_287": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2883943104, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1402.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2884414144, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1403.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2884415424, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1404.const", + "file_size": 1280 + }, + "onnx::MatMul_18435": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2884416704, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1405.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_207_bfp.out9_201_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2884887744, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1406.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.1/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_283_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2884887872, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1407.const", + "file_size": 128 + }, + "onnx::MatMul_18445_11_0_288": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2884888000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1408.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2885359040, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1409.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.1.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2885360320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1410.const", + "file_size": 1280 + }, + "onnx::MatMul_18446_11_0_289_12_67_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2885361600, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1411.const", + "file_size": 1868800 + }, + "onnx::MatMul_18446_11_0_289_12_67_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2887230400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1412.const", + "file_size": 1868800 + }, + "onnx::MatMul_18447_11_0_290": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2889099200, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1413.const", + "file_size": 1884160 + }, + "onnx::MatMul_18448_11_0_291": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2890983360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1414.const", + "file_size": 471040 + }, + "GroupNorm_36_wts_6_0_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1920 + ], + "size_in_bytes": 3840, + "op_tensor_size": 3840, + "offset": 2891454400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1415.const", + "file_size": 3840 + }, + "NhwcConv_37_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 6425600 + ], + "size_in_bytes": 6425600, + "op_tensor_size": 6425600, + "offset": 2891458240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1416.const", + "file_size": 6425600 + }, + "GroupNorm_37_wts_6_0_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2897883840, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1417.const", + "file_size": 2560 + }, + "NhwcConv_38_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "op_tensor_size": 4311040, + "offset": 2897886400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1418.const", + "file_size": 4311040 + }, + "NhwcConv_39_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 752640 + ], + "size_in_bytes": 752640, + "op_tensor_size": 752640, + "offset": 2902197440, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1419.const", + "file_size": 752640 + }, + "GroupNorm_38_wts_6_1_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2902950080, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1420.const", + "file_size": 2560 + }, + "onnx::MatMul_18455_11_0_292": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2902952640, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1421.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2903423680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1422.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2903424960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1423.const", + "file_size": 1280 + }, + "Attention_68_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2903426240, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1424.const", + "file_size": 471040 + }, + "Attention_68_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2903897280, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1425.const", + "file_size": 471040 + }, + "Attention_68_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2904368320, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1426.const", + "file_size": 471040 + }, + "onnx::MatMul_18466_11_0_293": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2904839360, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1427.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2905310400, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1428.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2905311680, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1429.const", + "file_size": 1280 + }, + "onnx::MatMul_18467": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2905312960, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1430.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_208_bfp.out9_204_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2905784000, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1431.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out2_3_288_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2905784128, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1432.const", + "file_size": 128 + }, + "onnx::MatMul_18477_11_0_294": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2905784256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1433.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2906255296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1434.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2906256576, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1435.const", + "file_size": 1280 + }, + "onnx::MatMul_18478_11_0_295_12_68_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2906257856, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1436.const", + "file_size": 1868800 + }, + "onnx::MatMul_18478_11_0_295_12_68_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2908126656, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1437.const", + "file_size": 1868800 + }, + "onnx::MatMul_18479_11_0_296": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2909995456, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1438.const", + "file_size": 1884160 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2911879616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1439.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2911880896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1440.const", + "file_size": 1280 + }, + "Attention_69_qkv_weight_key": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2911882176, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1441.const", + "file_size": 471040 + }, + "Attention_69_qkv_weight_query": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2912353216, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1442.const", + "file_size": 471040 + }, + "Attention_69_qkv_weight_value": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2912824256, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1443.const", + "file_size": 471040 + }, + "onnx::MatMul_18490_11_0_297": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2913295296, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1444.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2913766336, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1445.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2913767616, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1446.const", + "file_size": 1280 + }, + "onnx::MatMul_18491": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2913768896, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1447.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/to_q/MatMul_output_0.out2_6_209_bfp.out9_207_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2914239936, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1448.const", + "file_size": 128 + }, + "/unet/up_blocks.1/attentions.2/transformer_blocks.1/attn2/Reshape_3_output_0.out2_3_292_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2914240064, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1449.const", + "file_size": 128 + }, + "onnx::MatMul_18501_11_0_298": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2914240192, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1450.const", + "file_size": 471040 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2914711232, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1451.const", + "file_size": 1280 + }, + "unet.up_blocks.1.attentions.2.transformer_blocks.1.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2914712512, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1452.const", + "file_size": 1280 + }, + "onnx::MatMul_18502_11_0_299_12_69_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2914713792, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1453.const", + "file_size": 1868800 + }, + "onnx::MatMul_18502_11_0_299_12_69_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "op_tensor_size": 1868800, + "offset": 2916582592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1454.const", + "file_size": 1868800 + }, + "onnx::MatMul_18503_11_0_300": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "op_tensor_size": 1884160, + "offset": 2918451392, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1455.const", + "file_size": 1884160 + }, + "onnx::MatMul_18504_11_0_301": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "op_tensor_size": 471040, + "offset": 2920335552, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1456.const", + "file_size": 471040 + }, + "/unet/up_blocks.1/upsamplers.0/Resize.weights3_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 2920806592, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1457.const", + "file_size": 256 + }, + "NhwcConv_40_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4229120 + ], + "size_in_bytes": 4229120, + "op_tensor_size": 4229120, + "offset": 2920806848, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1458.const", + "file_size": 4229120 + }, + "GroupNorm_39_wts_6_0_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1920 + ], + "size_in_bytes": 3840, + "op_tensor_size": 3840, + "offset": 2925035968, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1459.const", + "file_size": 3840 + }, + "NhwcConv_41_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3171840 + ], + "size_in_bytes": 3171840, + "op_tensor_size": 3171840, + "offset": 2925039808, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1460.const", + "file_size": 3171840 + }, + "GroupNorm_40_wts_6_0_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2928211648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1461.const", + "file_size": 1280 + }, + "NhwcConv_42_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 2928212928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1462.const", + "file_size": 1057280 + }, + "NhwcConv_43_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 376320 + ], + "size_in_bytes": 376320, + "op_tensor_size": 376320, + "offset": 2929270208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1463.const", + "file_size": 376320 + }, + "GroupNorm_41_wts_6_0_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2929646528, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1464.const", + "file_size": 2560 + }, + "NhwcConv_44_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2114560 + ], + "size_in_bytes": 2114560, + "op_tensor_size": 2114560, + "offset": 2929649088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1465.const", + "file_size": 2114560 + }, + "GroupNorm_42_wts_6_0_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2931763648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1466.const", + "file_size": 1280 + }, + "NhwcConv_45_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 2931764928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1467.const", + "file_size": 1057280 + }, + "NhwcConv_46_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 250880 + ], + "size_in_bytes": 250880, + "op_tensor_size": 250880, + "offset": 2932822208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1468.const", + "file_size": 250880 + }, + "GroupNorm_43_wts_6_0_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "op_tensor_size": 2560, + "offset": 2933073088, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1469.const", + "file_size": 2560 + }, + "NhwcConv_47_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2114560 + ], + "size_in_bytes": 2114560, + "op_tensor_size": 2114560, + "offset": 2933075648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1470.const", + "file_size": 2114560 + }, + "GroupNorm_44_wts_6_0_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2935190208, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1471.const", + "file_size": 1280 + }, + "NhwcConv_48_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1057280 + ], + "size_in_bytes": 1057280, + "op_tensor_size": 1057280, + "offset": 2935191488, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1472.const", + "file_size": 1057280 + }, + "NhwcConv_49_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 250880 + ], + "size_in_bytes": 250880, + "op_tensor_size": 250880, + "offset": 2936248768, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1473.const", + "file_size": 250880 + }, + "GroupNorm_45_wts_6_0_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "op_tensor_size": 1280, + "offset": 2936499648, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1474.const", + "file_size": 1280 + }, + "GroupNorm_45.out7_45.out6_0_34_bfp.out7_45_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2936500928, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1475.const", + "file_size": 128 + }, + "NhwcConv_50_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 54400 + ], + "size_in_bytes": 54400, + "op_tensor_size": 54400, + "offset": 2936501056, + "file_name": ".cache/NhwcConv_0-unetconv_inConv_1476.const", + "file_size": 54400 + } + }, + "aux_info": {} } \ No newline at end of file