RuntimeError: Distributed package doesn't have NCCL built in (On Windows machine)

[justinjohn0306]

Any ideas on how to fix this?

[leng-yue]

This is probably a PyTorch / Lightning DDP environment problem, I believe there are some solutions on Stack overflow :)

[leng-yue]

Please don't send emails directly to my mailbox :) Using GitHub issues can help others to know and solve problems.

Original Email: Windows don't have NCCL if you can switch to gloo it might do the trick but I have no idea how to do that

To use gloo backend, you can edit the trainer config: configs/_base_/trainers/base.py

strategy=DDPStrategy(find_unused_parameters=True, process_group_backend="gloo"),

You can find it in pytorch-lightning's document

[justinjohn0306]

Yes! That did the trick. Thanks :)

[leng-yue]

Sounds good :)