Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

paraformer_large_offline triton运行bug 修复

yijinsheng2b747626c898
diff --git a/runtime/triton_gpu/Dockerfile/Dockerfile.server b/runtime/triton_gpu/Dockerfile/Dockerfile.server
index d03610c58..b9f9c4010 100644
--- a/runtime/triton_gpu/Dockerfile/Dockerfile.server
+++ b/runtime/triton_gpu/Dockerfile/Dockerfile.server
@@ -8,11 +8,18 @@ LABEL repository="tritonserver"
RUN apt-get update && apt-get -y install \
python3-dev \
cmake \
- libsndfile1
-
+ libsndfile1 \
+ language-pack-zh-hans
+ENV LANG="zh_CN.UTF-8"
+ENV LANGUAGE="zh_CN:zh:en_US:en"
# -i https://pypi.tuna.tsinghua.edu.cn/simple
-RUN pip3 install torch torchaudio
-RUN pip3 install kaldifeat pyyaml
+RUN pip3 config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+RUN pip3 install torch==2.4.1 torchaudio==2.4.1
+RUN pip3 install pyyaml
+
+##Refer to https://csukuangfj.github.io/kaldifeat/installation/from_wheels.html for installation.
+RUN pip3 install https://hf-mirror.com/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.5.dev20240914+cuda12.1.torch2.4.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+
# Dependency for client
RUN pip3 install soundfile grpcio-tools tritonclient
diff --git a/runtime/triton_gpu/README_paraformer_offline.md b/runtime/triton_gpu/README_paraformer_offline.md
index 48e889c29..ceaa5715b 100644
--- a/runtime/triton_gpu/README_paraformer_offline.md
+++ b/runtime/triton_gpu/README_paraformer_offline.md
@@ -10,6 +10,7 @@ pretrained_model_dir=$(pwd)/speech_paraformer-large_asr_nat-zh-cn-16k-common-voc
cp $pretrained_model_dir/am.mvn ./model_repo_paraformer_large_offline/feature_extractor/
cp $pretrained_model_dir/config.yaml ./model_repo_paraformer_large_offline/feature_extractor/
+cp $pretrained_model_dir/tokens.json ./model_repo_paraformer_large_offline/scoring/1/
# Refer here to get model.onnx (https://github.com/alibaba-damo-academy/FunASR/blob/main/funasr/export/README.md)
cp <exported_onnx_dir>/model.onnx ./model_repo_paraformer_large_offline/encoder/1/
@@ -33,6 +34,7 @@ model_repo_paraformer_large_offline/
`-- scoring
|-- 1
| `-- model.py
+ | -- tokens.json
`-- config.pbtxt
8 directories, 9 files
diff --git a/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py b/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
index d582e50ff..6f7e72064 100644
--- a/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
+++ b/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
@@ -74,8 +74,8 @@ class TritonPythonModel:
load lang_char.txt
"""
with open(str(vocab_file), "rb") as f:
- config = yaml.load(f, Loader=yaml.Loader)
- return config["token_list"]
+ vocab_list = json.load(f, encoding='utf-8')
+ return vocab_list
def execute(self, requests):
"""`execute` must be implemented in every Python model. `execute`
@@ -142,7 +142,7 @@ class TritonPythonModel:
]
responses = []
for i in range(total_seq):
- sents = np.array(hyps[i : i + 1])
+ sents = np.array(hyps[i: i + 1])
out0 = pb_utils.Tensor("OUTPUT0", sents.astype(self.out0_dtype))
inference_response = pb_utils.InferenceResponse(output_tensors=[out0])
responses.append(inference_response)
diff --git a/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt b/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
index 85082f44d..578fb303a 100644
--- a/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
+++ b/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
@@ -23,7 +23,7 @@ parameters [
},
{
key: "vocabulary",
- value: { string_value: "./model_repo_paraformer_large_offline/feature_extractor/config.yaml"}
+ value: { string_value: "./model_repo_paraformer_large_offline/scoring/1/tokens.json"}
},
{
key: "lm_path"