#5681 CUDA error: device-side assert triggeredCompile with TORCH_USE_CUDA_DSA to enable device-side assertions.,Traceback (mos

113.134* Posted at: 2 hours ago

CUDA error: device-side assert triggered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
,Traceback (most recent call last):
File "videotrans\process\tts_fun.py", line 127, in qwen3tts_fun
File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "K:\peiyinruanjian\win-pyvideotrans-v4.04-0704_3\_internal\qwen_tts\inference\qwen3_tts_model.py", line 603, in generate_voice_clone

talker_codes_list, _ = self.model.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "K:\peiyinruanjian\win-pyvideotrans-v4.04-0704_3\_internal\qwen_tts\core\models\modeling_qwen3_tts.py", line 2272, in generate

talker_result = self.talker.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "transformers\generation\utils.py", line 2566, in generate

result = decoding_method(

File "trans
......
nal\qwen_tts\core\models\modeling_qwen3_tts.py", line 1671, in forward

predictor_result = self.code_predictor.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "transformers\generation\utils.py", line 2566, in generate

result = decoding_method(

File "transformers\generation\utils.py", line 2831, in _sample

next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)

RuntimeError: CUDA error: device-side assert triggered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

Traceback (most recent call last):
File "videotrans\task\only_one.py", line 86, in run
File "videotrans\task\_stage_dubbing.py", line 23, in dubbing
File "videotrans\task\_stage_dubbing.py", line 102, in _tts
File "videotrans\tts\__init__.py", line 207, in run
File "videotrans\tts\_base.py", line 93, in run
File "videotrans\tts\_qwenttslocal.py", line 46, in _exec
File "videotrans\configure\base.py", line 272, in _new_process
videotrans.configure.excepts.VideoTransError: CUDA error: device-side assert triggered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
,Traceback (most recent call last):
File "videotrans\process\tts_fun.py", line 127, in qwen3tts_fun
File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "K:\peiyinruanjian\win-pyvideotrans-v4.04-0704_3\_internal\qwen_tts\inference\qwen3_tts_model.py", line 603, in generate_voice_clone

talker_codes_list, _ = self.model.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "K:\peiyinruanjian\win-pyvideotrans-v4.04-0704_3\_internal\qwen_tts\core\models\modeling_qwen3_tts.py", line 2272, in generate

talker_result = self.talker.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "transformers\generation\utils.py", line 2566, in generate

result = decoding_method(

File "transformers\generation\utils.py", line 2789, in _sample

outputs = model_forward(**model_inputs, return_dict=True)

File "torch\nn\modules\module.py", line 1751, in _wrapped_call_impl

return self._call_impl(*args, **kwargs)

File "torch\nn\modules\module.py", line 1762, in _call_impl

return forward_call(*args, **kwargs)

File "transformers\utils\generic.py", line 918, in wrapper

output = func(self, *args, **kwargs)

File "K:\peiyinruanjian\win-pyvideotrans-v4.04-0704_3\_internal\qwen_tts\core\models\modeling_qwen3_tts.py", line 1671, in forward

predictor_result = self.code_predictor.generate(

File "torch\utils\_contextlib.py", line 116, in decorate_context

return func(*args, **kwargs)

File "transformers\generation\utils.py", line 2566, in generate

result = decoding_method(

File "transformers\generation\utils.py", line 2831, in _sample

next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)

RuntimeError: CUDA error: device-side assert triggered
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

cfg=[TaskCfgVTT]当前工作模式: 翻译视频或音频 单视频模式
原始输入文件名: C:/Users/Administrator/Desktop/【看韩国搞笑短剧学韩语-中韩字幕】韩国职场上班第一天_哔哩哔哩_bilibili.mp4,
输出结果保存到文件夹: C:/Users/Administrator/Desktop/_video_out/【看韩国搞笑短剧学韩语-中韩字幕】韩国职场上班第一天_哔哩哔哩_bilibili-mp4,
临时文件夹: K:/peiyinruanjian/win-pyvideotrans-v4.04-0704_3/tmp/4176/1c02bdda04
已选 清理已存在
已选 启用CUDA加速
未选 降噪
语音识别:faster-whisper(本地内置), model_name: large-v3-turbo, 发音语言: 韩语, 断句方式:默认断句
翻译渠道:微软(免费),原始语言:韩语,目标语言:简体中文, 已选 发送完整字幕
配音渠道:Qwen3-TTS(本地内置), 角色:123.wav, 配音语言:简体中文, 未选 二次语音识别
音量:+0%, 语速:+0%, 未选 音频加速, 未选 视频慢速
未选 移除字幕间空隙, 已选 强制对齐字幕和声音
字幕: 不嵌入字幕
已选 分离人声与背景声, 已选 重新嵌入背景声, 背景音量0.8, 背景声音时长 短于 视频时长时: 拉长(降速播放),存在分离后的纯净人声文件,存在分离后的背景声音文件
代理地址:


=
system:Windows-10-10.0.19045-SP0
version:v4.04
frozen:True
language:zh
root_dir:K:/peiyinruanjian/win-pyvideotrans-v4.04-0704_3

1 Replies

这是由于 Qwen3-TTS 本地模型在尝试音色克隆(使用 123.wav 作为参考音频)时,CUDA 设备触发了内部断言(device-side assert),常见原因是参考音频格式不达标或模型输入数据出现越界/NaN。

解决方案:

  1. 检查参考音频 123.wav

    • 必须为 WAV 格式(.wav,采样率 16kHz单声道,时长 3–10 秒
    • 确保音频内容 只有清晰人声,无背景噪声、静音开头/结尾
    • 如果音频是其他格式或包含背景声,请先用专业工具(如 Audacity)裁剪并转换为正确格式。
    • 确认该文件已放置于 软件目录/f5-tts/ 文件夹内,并在 菜单 → TTS设置 → 设置参考音频 中正确填写了 123.wav#音频内的说话文本
  2. 尝试关闭 CUDA 加速临时测试

    • 在主界面取消勾选“CUDA加速”,改用 CPU 模式运行一次。
    • 如果 CPU 模式成功,说明是 CUDA 环境不稳定或显存不足导致的计算异常。
    • 如果 CPU 模式也失败,则问题基本确定在音频文件本身。
  3. 排查显存不足

    • Qwen3-TTS 本地模型 + large-v3-turbo 同时运行时,建议保证 显存 ≥ 8GB
    • 可在 高级选项 → 语音识别参数 中将 计算数据类型 改为 float16,或暂时将 ASR 模型换小(如 base)以降低显存占用。
  4. 回退其他配音渠道验证

    • 将配音渠道临时改为 Edge-TTS(免费免配置),选择一个中文角色(如 zh-CN-YunyangNeural),看是否能正常完成流程。
    • 如果能正常完成,则问题特定于 Qwen3-TTS + 克隆模式。
  5. 更换其他支持克隆的渠道

    • 如果 123.wav 确实符合要求但仍报错,可尝试改为 F5-TTS(本地内置)ZipVoice(本地内置),这两者对短参考音频更稳定。

注意:使用克隆角色时,高级选项 → 语音识别参数 中应将“最短语音持续毫秒”设为 3000、“最长语音持续秒数”设为 10,并勾选“Whisper预分割音频”和“合并过短字幕到相邻”,以保证截取的参考音频时长在 3–10 秒内。您的配置中未提到这些设置,请确认。

如果以上均无效,请提供 logs/ 目录下最新日志文件中的完整报错信息(复制约 30 行)以便进一步分析。

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. Dondate: ko-fi.com/jianchang512

Related resource