Training MusicGen on a Mac M4/64GB uses up all my memory, even with the small model

[rjjaret]

I'm trying to train MusicGen on a Mac M4, with 64GB of memory.

I'm using cpu and the small model. I'm only using the two sample audio files provided with audio craft, which I've converted to mono to simplify the test.

I've tried several things, but no matter what I do, the process is killed because all the memory gets used up. I've tried running it with float16 instead of 32. Seems I should have plenty of memory. Any thoughts? I've attached my config file.

Here's the command I'm using:

dora run solver=musicgen/musicgen_base_32khz model/lm/model_scale=small continue_from=//pretrained/facebook/musicgen-small conditioner=text2music device=cpu

musicgen_base_32khz.yaml

Thanks for any help you might offer.

[betweentwomidnights]

might want to try some other params in the dora run command?

" dataset.num_workers=1"
" dataset.valid.num_samples=1"
" dataset.batch_size=2"

I didn't even know it was possible to train using cpu. I don't have a mac or I'd help troubleshoot better. Might even be able to do batch_size=1 I'm not sure. If I recall it errors out when I tried that at one point.

I also notice in the config:

autocast: true

Maybe try flipping that to false.