Errors executing torchrun for train.py on Apple Silicon M2 Pro

[droptpackets]

When attempting to do fine tuning, I'm getting the following error in the output: RuntimeError: Distributed package doesn't have NCCL built in

Searching here indicates this is related to CUDA and other NVIDIA GPU related rendering.

So, I added the following snippet to train.py, which is supposed to force CPU only (same workaround used by this user in another meta-related repo: https://github.com/markasoftware/llama-cpu): torch.distributed.init_process_group("gloo")

Now the NCCL error goes away, but I get this error instead: AttributeError: module 'torch._C' has no attribute '_cuda_setDevice'

Again, I have no NVIDIA GPU or other such software on my system. I've tried a myriad of workarounds for my apple silicon, but haven't gotten very far.

Anything I'm missing here?

I'm running on python 3.10.9 with all requirements.txt entries installed as required via pip.