Using CPU instead of GPU

[Huba2004]

Hello!
I would like to train a glow_tts model in Hungarian on a MacBook Pro M1. CUDA GPU is not available, but it is very slow to train.

Is there any solution to make the training faster?

[erogol]

Nope, you need a GPU unfortunately.

[thefex]

There is a possibility to check if MPS (Apple Metal -> M1 GPU support) via torch.backends.is_mps_available() and convert .cuda() calls into .to(torch.device('mps')) however...
I tried and failed - this is a blocker: pytorch/pytorch#88415 - if that gets resolved that should be ~1 hour job.

Generally speaking it should be doable anyway ( but I am not python dev/experienced ML dev at all so it would take me too much time ) by replacing to specific MPS engine commands. Unfortunately MPS engine is not always as friendly as cuda - in CUDA pytorch simple .cuda() calls works.

You also need to make sure to update requirements.txt numba=0.55.1 to 0.55.2 as Apple M1 is supported in 0.55.2 (without that compile fails -> numba/numba#7951)

If someone experienced could pick this.. M1 machines/gpus are very strong and can yield better results that TESLA T4 etc..