funasr-server指定hub为ms也会从hf上下载模型

[ZhangQi-HUST]

🐛 Bug

1.3.25中 funasr-server指定hub为ms也会从hf上下载模型

BUG所在

funasr/bin/_server_app.py中142行

vllm_hub = app.state.hub if app.state.model_path else "hf"

和163行

"hub": app.state.hub if app.state.model_path else "hf",

逻辑有问题，只有model_path参数和hub参数都不为空时，hub才是ms，否则都是hf

忘修复

[LauraGPT]

Thanks for the precise report. You were right: the default Fun-ASR-Nano server path forced hub="hf" when model_path was not set, so funasr-server --model fun-asr-nano --hub ms could still try Hugging Face.

Fixed in #3324 and released in funasr==1.3.26:

python -m pip install -U "funasr==1.3.26"
funasr-server --model fun-asr-nano --hub ms --device cuda

What changed:

• vLLM loading now passes the requested server hub for the default FunAudioLLM/Fun-ASR-Nano-2512 model.
• AutoModel fallback loading now does the same if vLLM is unavailable.

Verification before release:

red first: the two new regression tests reproduced the old `hf` behavior
python3 -m pytest -q tests/test_server_app_openai_segments.py tests/test_docs_funasr_install_commands.py tests/test_collect_growth_metrics.py  # 60 passed
twine check dist/*  # wheel and sdist passed
public PyPI JSON: funasr 1.3.26 with expected hashes
fresh PyPI install: funasr.__version__ == "1.3.26"
GitHub Release: https://github.com/modelscope/FunASR/releases/tag/v1.3.26

I will leave this issue open briefly for your confirmation, but the fix is available on PyPI now.

[ZhangQi-HUST]

bug has been fixed, thanks

Thanks for the precise report. You were right: the default Fun-ASR-Nano server path forced hub="hf" when model_path was not set, so funasr-server --model fun-asr-nano --hub ms could still try Hugging Face.

Fixed in #3324 and released in funasr==1.3.26:

python -m pip install -U "funasr==1.3.26"
funasr-server --model fun-asr-nano --hub ms --device cuda
What changed:

• vLLM loading now passes the requested server hub for the default FunAudioLLM/Fun-ASR-Nano-2512 model.
• AutoModel fallback loading now does the same if vLLM is unavailable.

Verification before release:

red first: the two new regression tests reproduced the old `hf` behavior
python3 -m pytest -q tests/test_server_app_openai_segments.py tests/test_docs_funasr_install_commands.py tests/test_collect_growth_metrics.py  # 60 passed
twine check dist/*  # wheel and sdist passed
public PyPI JSON: funasr 1.3.26 with expected hashes
fresh PyPI install: funasr.__version__ == "1.3.26"
GitHub Release: https://github.com/modelscope/FunASR/releases/tag/v1.3.26

I will leave this issue open briefly for your confirmation, but the fix is available on PyPI now.

bug has been fixed.
Thanks a lot.