#4128 TaskCfgSTT(is_cuda=True, uuid='b2ec31ca8b', cache_folder='C:/pyvideotrans/tmp/6152/b2ec31ca8b', target_dir='C:/SOFT/pyvi

103.62* Posted at: 2 hours ago 👁9

用Qwen-ASR0.6B 语音识别文字,第一个任务用的tenvad,完成后再次添加语音选择silero VAD,执行时出现如下错误,任务中断后python进程没有终止,内容被大量占用,显存也被满占用。
语音识别阶段出错 [Qwen-ASR(本地)] CUDA error: an illegal memory access was encountered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

Traceback (most recent call last):
File "C:\pyvideotrans\videotrans\task\job.py", line 105, in run

trk.recogn()

File "C:\pyvideotrans\videotrans\task\_speech2text.py", line 159, in recogn

raw_subtitles = run(

File "C:\pyvideotrans\videotrans\recognition\__init__.py", line 259, in run

return QwenasrlocalRecogn(**kwargs).run()

File "C:\pyvideotrans\videotrans\recognition\_base.py", line 143, in run

res = self._exec()

File "C:\pyvideotrans\videotrans\recognition\_qwenasrlocal.py", line 48, in _exec

jsdata=self._new_process(callback=qwen3asr_fun,title=title,is_cuda=self.is_cuda,kwargs=kwargs)

File "C:\pyvideotrans\videotrans\configure\_base.py", line 289, in _new_process

raise RuntimeError(err)

RuntimeError: CUDA error: an illegal memory access was encountered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
TaskCfgSTT(is_cuda=True, uuid='b2ec31ca8b', cache_folder='C:/pyvideotrans/tmp/6152/b2ec31ca8b', target_dir='C:/SOFT/pyvideotrans/output/recogn', source_language=None, source_language_code=None, source_sub=None, source_wav=None, source_wav_output=None, target_language=None, target_language_code=None, target_sub='C:/SOFT/pyvideotrans/output/recogn/R_MIC_260405-124558.srt', target_wav=None, target_wav_output=None, name='K:/26404-411/R_MIC_260405-124558.mp3', noextname='R_MIC_260405-124558', basename='R_MIC_260405-124558.mp3', ext='mp3', dirname='K:/26404-411', shound_del_name=None, detect_language='zh-cn', recogn_type=2, model_name='0.6B', shibie_audio='C:/pyvideotrans/tmp/6152/b2ec31ca8b/R_MIC_260405-124558-1776467707.3235738.wav', remove_noise=False, enable_diariz=False, nums_diariz=0, rephrase=0, fix_punc=False)


=
system:Windows-10-10.0.22631-SP0
version:v3.99
frozen:False
language:zh
root_dir:C:/pyvideotrans
Python: 3.10.19 (main, Feb 12 2026, 00:37:08) [MSC v.1944 64 bit (AMD64)]

1 Replies

这是由于在切换 VAD 模型后,Qwen-ASR 本地识别过程中发生了 CUDA 非法内存访问错误,导致进程卡死并持续占用显存和内存。

解决方案:

  1. 强制终止残留进程:打开任务管理器(Ctrl+Shift+Esc),找到并结束所有残留的 Python 进程。
  2. 检查 CUDA 环境:确保已正确安装 CUDA 12.8 和 cuDNN 9.x,并且显卡驱动为最新版本。在 CMD 中运行 nvcc -Vnvidia-smi 进行验证。
  3. 调整高级设置以降低显存压力:在软件中点击 菜单 -> 工具/选项 -> 高级选项,在“语音识别参数”区域进行如下修改:

    • CUDA数据类型:从 float32 改为 float16int8
    • 识别准确度beam_size:从 5 改为 1
    • 识别准确度best_of:从 5 改为 1
    • 启用上下文感知:确保为 false(不勾选)。
  4. 使用更稳定的 VAD 设置:建议在切换 VAD 模型后,先处理一个短小的测试文件,观察是否稳定。如果问题依旧,可以尝试换回 ten-vad 或使用默认 VAD。
  5. 处理任务前清理缓存:在软件主界面左上角勾选 清理已生成,确保每次处理都使用新的参数设置,避免缓存干扰。

如果问题持续存在,请提供软件根目录下 logs 文件夹中最新日志文件的底部约30行内容,以便进一步分析。

请查阅相关文档:

Post Your Reply
Open source and free maintenance is not easy. If this project is helpful to you, please consider making a small donation to help the project continue to maintain and update.

Related resource