applied PR #1152 to other recipes

egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx-zh.py

@@ -613,7 +613,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx.py

@@ -608,7 +608,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless/export-onnx.py

@@ -508,7 +508,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless3/export-onnx.py

@@ -518,7 +518,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py

@@ -63,8 +63,8 @@ import sentencepiece as spm
 import torch
 import torch.nn as nn
 from conformer import Conformer
-from onnxruntime.quantization import QuantType, quantize_dynamic
 from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
 from scaling_converter import convert_scaled_to_non_scaled
 from train import add_model_arguments, get_params, get_transducer_model
 
@@ -662,7 +662,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx.py

@@ -60,8 +60,8 @@ import sentencepiece as spm
 import torch
 import torch.nn as nn
 from conformer import Conformer
-from onnxruntime.quantization import QuantType, quantize_dynamic
 from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
 from scaling_converter import convert_scaled_to_non_scaled
 from train import add_model_arguments, get_params, get_transducer_model
 
@@ -586,7 +586,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py

@@ -661,7 +661,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py

@@ -652,7 +652,7 @@ def main():
     quantize_dynamic(
         model_input=decoder_filename,
         model_output=decoder_filename_int8,
-        op_types_to_quantize=["MatMul"],
+        op_types_to_quantize=["MatMul", "Gather"],
         weight_type=QuantType.QInt8,
     )