RuntimeError: Distributed package doesn't have NCCL built in

[mdicillo]

Hello! First of all, great work on this! I've been playing with it all day and exploring its potential. I wanted to start training it on some original music and have run into an issue.

After starting "trainer.py" with the applicable flags, it will run for a little bit and then throw up the error:

RuntimeError: Distributed package doesn't have NCCL built in

This is on a Windows 11 machine and from some initial research, it appears that Windows doesn't support NCCL? I don't really know what NCCL is or if it's necessary, so I can't speak to this.

This is running on an RTX 3060 with CUDA available and running...

Any tips or information would be greatly appreciated!

Again, thanks for the awesome work!

[mdicillo]

I just solved this...I had to edit the trainer.py script to use the "gloo" backend:

Added this to the import section:
from pytorch_lightning.strategies import DDPStrategy

Replaced this at line 849:
strategy=DDPStrategy(process_group_backend="gloo"),

And now the trainer runs!

I hope this helps someone in the future...