From db71b0302651d0fd6d0e1c742591f35f2ab224ac Mon Sep 17 00:00:00 2001 From: Wei Kang Date: Thu, 29 Jun 2023 16:48:59 +0800 Subject: [PATCH] Support int8 quantization in decoder (#1152) --- egs/librispeech/ASR/zipformer/export-onnx-streaming.py | 2 +- egs/librispeech/ASR/zipformer/export-onnx.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py index ff3e46433..3eb06f68c 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py +++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py @@ -757,7 +757,7 @@ 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, ) diff --git a/egs/librispeech/ASR/zipformer/export-onnx.py b/egs/librispeech/ASR/zipformer/export-onnx.py index 1bc10c896..724fdd2a6 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx.py +++ b/egs/librispeech/ASR/zipformer/export-onnx.py @@ -602,7 +602,7 @@ 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, )