torch.distributed not supported on Windows, but throws error for non-distributed training

[sgbaird]

Hi,

Can't seem to train:

runfile('main.py','--mode train --config-yml configs/s2ef/200k/cgcnn/cgcnn.yml')
Traceback (most recent call last):

  File "main.py", line 15, in <module>
    from ocpmodels.common import distutils

  File "C:\Users\sterg\Documents\GitHub\sparks-baird\ocp\ocpmodels\common\distutils.py", line 98, in <module>
    def broadcast(tensor, src, group=dist.group.WORLD, async_op=False):

AttributeError: module 'torch.distributed' has no attribute 'group'

Possibly because torch.distributed isn't supported on Windows(?).

There may have been some progress with Windows support, however, it seemed like the command I used main.py --mode train --config-yml configs/s2ef/200k/cgcnn/cgcnn.yml shouldn't require torch.distributed, correct? Is there something basic I'm missing here?

Sterling

[mshuaibii]

We still rely on torch.distributed for other calls throughout the repo - even if not being used for distributed training. I'd recommend using WSL if working on a Windows machine (looks like you are #207) as we haven't supported Windows at this time.

@anuroopsriram if you have anything else to add to this?

[sgbaird]

Seems that Windows is supported now (at least in some sense) pytorch/pytorch#37068 (comment)