#7943 CUDA error: device-side assert triggeredCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

The CUDA error: device-side assert triggered is a GPU-side data validity error (in ChatterBox TTS this happens inside s3gen.embed_ref, i.e. the reference audio slice or token index fed to the GPU is out of the model's valid range). With the clone voice role the reference audio is dynamically cut out of your original file per subtitle line, so an invalid/too-short slice triggers it. Once this assert fires, the CUDA context is destroyed — every following GPU call fails too, so a simple retry cannot recover.

Steps to fix (in order):

1. Fully restart the software. Do not just click retry — the CUDA context must be rebuilt.

2. Fix reference audio length (main cause with clone):

   • Go to Menu -> Tools/Options -> Advanced Options -> ASR Settings ([语音识别参数])
   • VAD:max speech duration(s): set 6–10
   • VAD:min speech duration(ms): set 3000–4000 (must be ≥3000 for voice cloning — shorter slices cannot be used as reference audio)
   • Note: max duration (s) × 1000 must be greater than min duration (ms), otherwise recognition gets corrupted.

3. Check VRAM:

   • Close other GPU-consuming programs; ChatterBox on GPU needs roughly 10GB+ free VRAM.
   • In Advanced Options -> ASR Settings, set Compute data type to float16 (GPU-friendly) or int8 (lowest memory).
   • In Advanced Options -> General, keep Number of GPU tasks = 1.
4. Clean up the file path. Your input is
   C:/Users/Jonathan/Music/_video_out/1_acor34_(Vocals).flac
   Parentheses and other special characters are poorly handled by the underlying ffmpeg slicing. Rename it to plain ASCII letters/digits/underscore, e.g. 1_acor34_vocals.flac, and move it to a shallow path such as D:/vt/. (Also note the odd double nested output folder _video_out/_video_out/ — clear the output folder and let the software generate it.)

5. Verify your installation is clean. The log reports version:v4.15 but root_dir:C:/Users/Jonathan/Music/win-pyvideotrans-v4.08-0802, i.e. a mixed/partially overwritten installation:

   • Download the full package (about 2.7GB) and overwrite the whole folder, then overwrite the latest patch package again (all files and the _internal and videotrans folders, not only sp.exe).
   • If version-mismatch errors persist, search inside _internal for every __pycache__ folder and delete them before re-copying.
6. To isolate the fault: run once with Edge-TTS (free) as the dubbing channel. If it succeeds, the pipeline and your media are fine and the problem is specific to the ChatterBox clone reference audio; if it fails too, the problem is in the earlier stages (ASR/audio slicing).

Also confirm ChatterBox extra files are present: features.msgpack and weights.npz extracted into models/pkuser_home; if the model download was incomplete, the same kind of crash can occur.

If none of the above resolves it, please provide the latest log file from the logs/ folder (bottom ~30 lines) of a clean run after restarting, together with whether the step-6 Edge-TTS test succeeded.

请查阅相关文档：

• Advanced Options
• ChatterBox (built-in)
• All dubbing channels
• Voice cloning
• Optimize sentence segmentation / VAD
• FAQ