GPU using from cli

[jooni22]

Is it possible to use subsai from console (bash) to use GPU instead of CPU?

[absadiki]

Yes it is possible if the model supports GPU,
Actually you can use any parameter in the model-configs argument
For openai/whisper for example, you can check the list of all parameters here
and you can use it in from the CLI as follows :

subsai ./assets/test1.mp4 --model openai/whisper --model-configs '{"model_type": "small", "device": "cuda"}' --format srt

Hope this helps ?

[jooni22]

Yep, that working as well. I had issue with Nvidia and after fixing the subsai automatically set cuda if model supports GPU. Many thanks

[absadiki]

you're welcome :)

[gandarufu]

Which version of Cuda and cuDNN do you suggest? I currently tried 12.1 and i can't get subsAI to run on my gpu.

[s-h-a-d-o-w]

It's been 2 years but... this is still a problem and in case somebody else stumbles across it, this is how I made it work:

1. Wipe all globally installed packages. (Don't know whether that's relevant if you're using a virtual environment, I rarely work with Python.)
2. Optional: Create a virtual environment

python -m venv subsai
source subsai/bin/activate  # On Windows: subsai\Scripts\activate

3. Install subsai
4. Check CUDA version: nvcc --version
5. Get the appropriate installer command from the PyTorch website and run it.
6. Run subsai.

(I then got numerous warnings because different things rely on different, incompatible dependency versions but... it works.)

[absadiki]

Thanks @s-h-a-d-o-w