PyTorch MPS is buggy. Even with the stable build. Something with cuda is far better imo. Personally I use a mbp 14’ with the M1 Pro base model for literally everything and then I have a desktop (had one cuz I play games, just upgraded the gpu to a cheap 3090 I found online, works like a charm for 99% of work loads when it comes to training something.

For the 1% when I do not have enough compute I use my universities cluster or compute Canada for distributed training.