Randomly Crashing without Error

Environment

• OS: Windows 11
• GPU: RTX 3060 12GB
• funasr version: 1.3.1
• Python 3.10.9

Description

The application appears to freeze or terminate after generating some audios.
After some progress output, the console stops and eventually returns:

Drücken Sie eine beliebige Taste . . .

The issue happens consistently on my system.

Relevant Logs

funasr version: 1.3.1

Running on device: cuda, dtype: bfloat16
Warning: torch.compile disabled - triton is not installed

Loaded VoxCPM2Model

Warm up VoxCPMModel...

100%|██████████████████████████████████████████████████████████████████████████████████| 10/10 [00:02<00:00, 3.93it/s]

17%|█████████████▍                                                                   | 77/466 [00:17<01:28, 4.41it/s]
17%|██████████████                                                                   | 81/466 [00:19<01:31, 4.20it/s]
15%|████████████▏                                                                    | 70/466 [00:16<01:33, 4.23it/s]

Drücken Sie eine beliebige Taste . . .

Additional Notes

• CUDA is detected correctly.
• The model loads successfully before the freeze/crash.
• torch.compile is disabled because Triton is not installed.
• The issue occurs during the warmup/inference stage, not during model download.

Question

Is this a known issue on Windows or related to:

• Triton not being installed?
• bfloat16 support on RTX 3060?
• VoxCPM warmup/inference on Windows?

Any help or debugging suggestions would be appreciated.

Full Log:

funasr version: 1.3.1.
Downloading Model from https://www.modelscope.cn to directory: C:\Users\alexa\.cache\modelscope\hub\models\iic\SenseVoiceSmall
WARNING:root:trust_remote_code: False
* Running on local URL:  http://0.0.0.0:8808
* To create a public link, set `share=True` in `launch()`.
rtf_avg: 0.046: 100%|████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  2.19it/s]
Fetching 9 files: 100%|██████████████████████████████████████████████████████████████████████████| 9/9 [00:00<?, ?it/s]
Download complete: : 0.00B [00:00, ?B/s]              voxcpm_model_path: C:\Users\alexa\.cache\huggingface\hub\models--openbmb--VoxCPM2\snapshots\bffb3df5a29440629464e5e839f4d214c8714c3d, zipenhancer_model_path: iic/speech_zipenhancer_ans_multiloss_16k_base, enable_denoiser: True
Download complete: : 0.00B [00:00, ?B/s]
Loading AudioVAE from pytorch: C:\Users\alexa\.cache\huggingface\hub\models--openbmb--VoxCPM2\snapshots\bffb3df5a29440629464e5e839f4d214c8714c3d\audiovae.pth
Running on device: cuda, dtype: bfloat16
Loading model from safetensors: C:\Users\alexa\.cache\huggingface\hub\models--openbmb--VoxCPM2\snapshots\bffb3df5a29440629464e5e839f4d214c8714c3d\model.safetensors
Warning: torch.compile disabled - triton is not installed
Loaded VoxCPM2Model
Downloading Model from https://www.modelscope.cn to directory: C:\Users\alexa\.cache\modelscope\hub\models\iic\speech_zipenhancer_ans_multiloss_16k_base
2026-05-10 20:27:09,750 - modelscope - INFO - initiate model from C:\Users\alexa\.cache\modelscope\hub\models\iic\speech_zipenhancer_ans_multiloss_16k_base
2026-05-10 20:27:09,751 - modelscope - INFO - initiate model from location C:\Users\alexa\.cache\modelscope\hub\models\iic\speech_zipenhancer_ans_multiloss_16k_base.
2026-05-10 20:27:09,753 - modelscope - INFO - initialize model from C:\Users\alexa\.cache\modelscope\hub\models\iic\speech_zipenhancer_ans_multiloss_16k_base
2026-05-10 20:27:09,979 - modelscope - WARNING - No preprocessor field found in cfg.
2026-05-10 20:27:09,980 - modelscope - WARNING - No val key and type key found in preprocessor domain of configuration.json file.
2026-05-10 20:27:09,980 - modelscope - WARNING - Cannot find available config to build preprocessor at mode inference, current config: {'model_dir': 'C:\\Users\\alexa\\.cache\\modelscope\\hub\\models\\iic\\speech_zipenhancer_ans_multiloss_16k_base'}. trying to build by task and model information.
2026-05-10 20:27:09,981 - modelscope - INFO - No preprocessor key ('speech_zipenhancer_ans_multiloss_16k_base', 'acoustic-noise-suppression') found in PREPROCESSOR_MAP, skip building preprocessor. If the pipeline runs normally, please ignore this log.
Warm up VoxCPMModel...
100%|██████████████████████████████████████████████████████████████████████████████████| 10/10 [00:02<00:00,  3.93it/s]
 17%|█████████████▍                                                                   | 77/466 [00:17<01:28,  4.41it/s]
 17%|██████████████                                                                   | 81/466 [00:19<01:31,  4.20it/s]
 15%|████████████▏                                                                    | 70/466 [00:16<01:33,  4.23it/s]
Drücken Sie eine beliebige Taste . . .

Start by reproducing the VoxCPM warmup/inference run on Windows 11 with the RTX 3060 and compare the bfloat16, Triton-disabled configuration shown in the logs. Investigate the failure after the warmup progress output; done means identifying a reproducible cause or workaround and documenting the relevant environment requirements.