Upload LiteRT FP16 multi-sig bundle

README.md

@@ -63,15 +63,23 @@ Each signature has the same I/O shape contract:
 ```
 inputs:
   audio_signal : float32 [1, 128, T_mel]   # log-mel features (NeMo preproc)
-  length       : int64   [1]                # actual mel frames used (≤ T_mel)
+  length       : int32   [1]                # actual mel frames used (≤ T_mel)
 outputs:
   encoded         : float32 [1, 1024, T_enc]  # T_enc = (T_mel - 4) // 8
-  encoded_lengths : int64   [1]
+  encoded_lengths : int32   [1]
 ```
 
 Pick the smallest bucket that fits your input; pad shorter inputs with zeros
 and pass the true length.
 
+**Why int32, not int64.** LiteRT's GPU/NPU delegates (LiteRT-CL / OpenCL,
+NPU accelerator) reject int64 tensors entirely. With int64 length, every
+internal CAST node touching it falls back to CPU and `CompiledModel.create()`
+fails outright on Android with the GPU backend. This bundle is exported with
+int32 length end-to-end (input → internal mask arange/comparisons → output
+`encoded_lengths`). int32 covers >2 billion mel frames (~5 hours), so no
+practical range loss.
+
 ## Decoder + joint contract
 
 ```

encoder_multisig.tflite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a97075644590cedce95a53083c876f56dce22d2e1e5807bc4ca2d6879f6183c8
-size 1249026196
+oid sha256:bd24fd99bc73f92b5f9eab066e5e1bed907e0e2d733cb3472a4f0512c076e002
+size 1249003688

manifest.json

@@ -27,6 +27,8 @@
     "n_layers": 24,
     "n_heads": 8,
     "feat_in": 128,
+    "attention_mode": "rel_pos",
+    "att_context_size": null,
     "buckets": [
       {
         "n_mel_frames": 300,
@@ -171,8 +173,8 @@
         "graph": "encoder",
         "source_artifact": "encoder_dynamicT.pt2",
         "output_artifact": "encoder_multisig.tflite",
-        "size_mb": 1191.16,
-        "convert_seconds": 402.16,
+        "size_mb": 1191.14,
+        "convert_seconds": 367.97,
         "quant": "fp16",
         "multisig": true,
         "signatures": [
@@ -184,33 +186,33 @@
         "parity_per_signature": {
           "forward_T300": {
             "ok": true,
-            "max_abs_diff": 0.0033329054713249207,
+            "max_abs_diff": 0.009477382525801659,
             "per_output_diffs": [
-              0.0033329054713249207,
+              0.009477382525801659,
               0.0
             ]
           },
           "forward_T500": {
             "ok": true,
-            "max_abs_diff": 0.006780040450394154,
+            "max_abs_diff": 0.0061398837715387344,
             "per_output_diffs": [
-              0.006780040450394154,
+              0.0061398837715387344,
               0.0
             ]
           },
           "forward_T700": {
             "ok": true,
-            "max_abs_diff": 0.0005690590478479862,
+            "max_abs_diff": 0.001271696761250496,
             "per_output_diffs": [
-              0.0005690590478479862,
+              0.001271696761250496,
               0.0
             ]
           },
           "forward_T1500": {
             "ok": true,
-            "max_abs_diff": 0.003892328590154648,
+            "max_abs_diff": 0.004102766513824463,
             "per_output_diffs": [
-              0.003892328590154648,
+              0.004102766513824463,
               0.0
             ]
           }
@@ -221,7 +223,7 @@
         "source_artifact": "decoder_step.pt2",
         "output_artifact": "decoder_step.tflite",
         "size_mb": 22.55,
-        "convert_seconds": 3.81,
+        "convert_seconds": 2.72,
         "quant": "fp16",
         "torch_output_shapes": [
           [
@@ -313,7 +315,7 @@
         "source_artifact": "joint_step.pt2",
         "output_artifact": "joint_step.tflite",
         "size_mb": 12.08,
-        "convert_seconds": 1.13,
+        "convert_seconds": 1.08,
         "quant": "fp16",
         "torch_output_shapes": [
           [
@@ -325,9 +327,9 @@
         ],
         "parity": {
           "ok": true,
-          "max_abs_diff": 0.275390625,
+          "max_abs_diff": 0.33984375,
           "per_output_diffs": [
-            0.275390625
+            0.33984375
           ],
           "tflite_output_shapes": [
             [