vllm realtime_ws报错

[yourFinger]

Notice: In order to resolve issues efficiently, please follow the template and include reproducible details.
（注意：为了更加高效率解决您遇到的问题，请按照模板提问，补充可复现细节。）

🐛 Bug

To Reproduce

Steps to reproduce the behavior. Always include the exact command you ran.

1. Install with: ...
2. Run: ...
3. See error: ...

Code sample

#!bin/bash

CUDA_VISIBLE_DEVICES=0
nohup python examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py --port 10095 --language 中文 --hotword-file hotword_list > serve_realtime_ws.log 2>&1 &

[serve_realtime_ws.log](https://github.com/user-attachments/files/28372320/serve_realtime_ws.log)

[serve_realtime_ws.log](https://github.com/user-attachments/files/28372328/serve_realtime_ws.log)

Expected behavior

Error logs

Environment

• OS:
• Python version: 3.12
• FunASR version: 1.3.7
• ModelScope version:
• PyTorch / torchaudio version:
• Install method (pip, source, Docker):
• Device (cuda, cpu, mps):
• GPU model: RTX 4090
• CUDA/cuDNN version: 13.2
• Docker image tag, if used:

Audio details

If the audio cannot be shared, please describe:

• Duration:
• Sample rate:
• Format:
• Language/dialect:
• Speaker count:
• Background noise/music:

[yourFinger]

serve_realtime_ws.log

[qiulang]

你这个错估计就是我报的 #2950

[LauraGPT]

This is the same issue as #2950 — repetition_penalty is incompatible with enable_prompt_embeds=True in vLLM.

Already fixed on main branch (commit b8d736e): removed repetition_penalty from both serve_realtime_ws.py and serve_vllm.py, replaced with post-processing repetition truncation.

Temporary fix until next PyPI release: remove the repetition_penalty=1.3 line from your local serve_realtime_ws.py:

# Line ~321: delete this line
# repetition_penalty=1.3,

Or pull latest from GitHub:

git pull origin main

[LauraGPT]

This bug is fixed in v1.3.8 (just published today).

The root cause was repetition_penalty=1.3 in the vLLM sampling params, which is incompatible with enable_prompt_embeds=True mode. vLLM's repetition penalty implementation does a scatter operation on prompt token IDs, but in prompt-embeds mode there are no token IDs — causing the CUDA assert.

Fix: Upgrade to v1.3.8:

pip install --upgrade funasr

The fix removes repetition_penalty from vLLM sampling and replaces it with post-processing repetition detection (truncate_repetition() + detect_and_fix_hallucination()), which achieves the same effect without the CUDA incompatibility.

[LauraGPT]

Update: v1.3.9 is now released with this fix. Upgrade with:

pip install --upgrade funasr

This version also publishes a wheel for faster installation.