[Bug]: Issue when running Qwen3 TTS example and let vllm omni to download weight on the fly

[tjtanaa]

Your current environment

The output of python collect_env.py

Your output of `python collect_env.py` here

Your code version

The commit id or version of vllm

The commit id or version of vllm-omni

🐛 Describe the bug

Without downloading the weights before hand, I run the example python end2end.py --query-type CustomVoice and python end2end.py --query-type VoiceDesign . Both cases I encounter this error. It seems the files are not downloaded completely before it is trying to load the preprocessor_config.json . I have to predownload the weights by hf download Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign

(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749]   File "/usr/local/lib/python3.12/dist-packages/transformers/models/auto/
feature_extraction_auto.py", line 358, in from_pretrained                                                                                              
(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749]     config_dict, _ = FeatureExtractionMixin.get_feature_extractor_dict(pr
etrained_model_name_or_path, **kwargs)                                                                                                                 
(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749]                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^                                                                                                                 
(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749]   File "/usr/local/lib/python3.12/dist-packages/transformers/feature_extr
action_utils.py", line 536, in get_feature_extractor_dict                                                                                              
(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749]     raise OSError(                                                       
(Worker pid=21493) [Stage-0] ERROR 01-22 15:11:20 [multiproc_executor.py:749] OSError: Can't load feature extractor for '/app/model/models--Qwen--Qwen3
-TTS-12Hz-1.7B-VoiceDesign/snapshots/385f44a7d86fa76e503b8633f22a5196b999e53b/speech_tokenizer'. If you were trying to load it from 'https://huggingfac
e.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure '/app/model/models--Qwen--Qwen3-TTS-12Hz-1.7B-VoiceDe
sign/snapshots/385f44a7d86fa76e503b8633f22a5196b999e53b/speech_tokenizer' is the correct path to a directory containing a preprocessor_config.json file
(Worker pid=21493) [Stage-0] INFO 01-22 15:11:20 [multiproc_executor.py:707] Parent process exited, terminating worker                                 
[rank0]:[W122 15:11:21.428916169 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can lea
k resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())  

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[NickLucche]

this is actually working on my side, although some of the dependencies weren't installed by default: onnxruntime (why is ort even needed?) + sox

[Gaohan123]

The issue is fixed in PR #903. The dependencies will be fixed in next PR.

[tjtanaa]

@Gaohan123 Regarding to onnruntime dependencies, different platform has different ways to install onnxruntime e.g. on ROCm

pip uninstall onnxruntime # we must remove this first before installing onnxruntime-rocm
pip install onnxruntime-rocm sox

So, we will need a platform dependent package installation step as well.

[gcanlin]

Dependencies: #909

[tjtanaa]

Dependencies: #909

I added my reply in the PR as a comment.

[Gaohan123]

#981 fix dependencies. This issue will be closed.