Get an error in ffmpeg_audio() func when adding --mdx_device cuda:1

[xueli12]

Why do I get an error when I add --mdx_device cuda:1, but it works fine without it?

Faster Whisper execution command:

./Faster-Whisper-XXL/Faster-Whisper-XXL/faster-whisper-xxl -m large-v2-japanese-5k-steps --model_dir models --verbose true videos/4106665-2.wav -l ja -d cuda:1 --output_format srt --skip -o /working/subtitles --vad_filter true --vad_threshold 0.22 --vad_method silero_v4_fw --ff_mdx_kim2 --one_word 0 --sentence --max_line_width 18 --max_line_count 1 --max_comma 5 --max_comma_cent 50 --vad_device cuda:1 --mdx_device cuda:1

2025-02-07 15:18:35,817 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Standalone Faster-Whisper-XXL r245.2 running on: CUDA
2025-02-07 15:18:35,817 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Standalone Faster-Whisper-XXL r245.2 running on: CUDA
2025-02-07 15:18:35,819 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Number of visible GPU devices: 2
2025-02-07 15:18:35,819 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Number of visible GPU devices: 2
2025-02-07 15:18:35,822 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Supported compute types by GPU: {'int8', 'float32', 'int8_float16', 'int8_float32', 'float16'}
2025-02-07 15:18:35,822 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Supported compute types by GPU: {'int8', 'float32', 'int8_float16', 'int8_float32', 'float16'}
2025-02-07 15:18:35,824 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Audio filtering is in progress...
2025-02-07 15:18:35,824 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Output: Audio filtering is in progress...
2025-02-07 15:18:35,852 - ERROR - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Error: Error: Some error in ffmpeg_audio() func.
2025-02-07 15:18:35,852 - ERROR - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Error: Error: Some error in ffmpeg_audio() func.
2025-02-07 15:18:36,228 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Faster Whisper return value: 130
2025-02-07 15:18:36,228 - INFO - [Task 1] [File videos/[4106665-2.wav](http://4106665-2.wav/)] Faster Whisper return value: 130

[Purfview]

What is "cuda:1" device and why you use it?
Why there is everything repeated twice in console?

[xueli12]

What is "cuda:1" device and why you use it? Why there is everything repeated twice in console?

Actually, I am not sure which GPU cuda:1 refers to. The reason for using it is to leverage multithreading to fully utilize both GPUs and make them work simultaneously.
As for the repeated log entries, I am also unsure of the cause. It might be a bug in the console.

!nvidia-smi
Fri Feb 7 16:15:28 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03 Driver Version: 560.35.03 CUDA Version: 12.6 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 Tesla T4 Off | 00000000:00:04.0 Off | 0 |
| N/A 43C P8 9W / 70W | 1MiB / 15360MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 1 Tesla T4 Off | 00000000:00:05.0 Off | 0 |
| N/A 44C P8 10W / 70W | 1MiB / 15360MiB | 0% Default |
| | | N/A |

[xueli12]

What is "cuda:1" device and why you use it? Why there is everything repeated twice in console?

What puzzles me is that using these parameters together results in an error:
-d cuda:1 --vad_method silero_v4_fw --vad_device cuda:1 --ff_mdx_kim2 --mdx_device cuda:1
However, it runs fine without --mdx_device cuda:1, i.e., with:
-d cuda:1 --vad_method silero_v4_fw --vad_device cuda:1 --ff_mdx_kim2

[Purfview]

The reason for using it is to leverage multithreading to fully utilize both GPUs and make them work simultaneously.

And how you do that?

--vad_method silero_v4_fw --vad_device cuda:1

Silero is cpu only.

What puzzles me is that using these parameters together results in an error:
-d cuda:1 --vad_method silero_v4_fw --vad_device cuda:1 --ff_mdx_kim2 --mdx_device cuda:1

Probably a bug. That's on kaggle? How much RAM it offers for free?

[xueli12]

Probably a bug. That's on kaggle? How much RAM it offers for free?

Yes, 30GB

[xueli12]

--vad_method silero_v4_fw --vad_device cuda:1

Silero is cpu only.

Thank you for pointing out the error. By the way, does adding the parameter --vad_device cuda:1 have any impact?

[Purfview]

I'll test it in kaggle to see what's up with "--mdx_device cuda:1".
As both GPU devices there are same then there's no point to set it to the second one.

By the way, does adding the parameter --vad_device cuda:1 have any impact?

It has no effect with silero.

[Purfview]

Fixed in r245.3

[xueli12]

Fixed in r245.3

r245.4 linux
I still encountered an error when using --voc_device cuda:1, but there was no issue when using --voc_device cuda: 
Loading model...
[Task 1] [File videos/222.wav] Starting processing file
[Task 1] [File videos/222.wav] Using device: cuda:1
[Task 1] [File videos/222.wav] Transcribing in progress...
[Task 1] [File videos/222.wav] Faster Whisper command: ./Faster-Whisper-XXL/Faster-Whisper-XXL/faster-whisper-xxl -m large-v2-japanese-5k-steps --model_dir models --print_progress videos/222.wav -l ja -d cuda:1 --output_format srt --skip -o /kaggle/working/subtitles --vad_filter true --vad_threshold 0.22 --vad_method pyannote_v3 --vad_device cuda:1 --one_word 0 --sentence --max_line_width 18 --max_line_count 1 --max_comma 5 --max_comma_cent 50 --ff_vocal_extract mdx_kim2 --voc_device cuda:1
2025-07-04 17:54:10,274 - INFO - [Task 1] [File videos/222.wav] Output: Standalone Faster-Whisper-XXL r245.4 running on: CUDA
2025-07-04 17:54:10,278 - INFO - [Task 1] [File videos/222.wav] Output: Starting to process: videos/222.wav
2025-07-04 17:54:10,284 - INFO - [Task 1] [File videos/222.wav] Output: Audio filtering is in progress...
2025-07-04 17:54:10,308 - INFO - [Task 1] [File videos/222.wav] Output: Traceback (most recent call last):
2025-07-04 17:54:10,312 - INFO - [Task 1] [File videos/222.wav] Output: File "main.py", line 176, in ffmpeg_audio
2025-07-04 17:54:10,312 - INFO - [Task 1] [File videos/222.wav] Output: File "main.py", line 176, in ffmpeg_audio
2025-07-04 17:54:10,316 - INFO - [Task 1] [File videos/222.wav] Output: File "faster_whisper/mdx_infer.py", line 117, in init
2025-07-04 17:54:10,319 - ERROR - [Task 1] [File videos/222.wav] Error: NameError: name 'device' is not defined
2025-07-04 17:54:10,644 - INFO - [Task 1] [File videos/222.wav] Faster Whisper returned code: 130

[Purfview]

But it should have been fixed in r245.3... maybe it regressed in r245.4!?

I'll check it again.

[Purfview]

Do you run two tasks at the same time there?

[xueli12]

Do you run two tasks at the same time there?

I've tried running it both separately and simultaneously — the result is the same.