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

[minar09]

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

Still getting this error after installing latest pytorch-nightly and other stuff, when I tried to run the imagenet example script.

[TimZaman]

Same here; MacOS, Python 3.7 through pip. pytorch version 1.0.0.

> dir(torch.distributed)
['__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '__package__', '__path__', '__spec__', 'deprecated', 'is_available', 'torch']

[soumith]

the OSX binary of PyTorch does not ship with distributed support. You can check this with: torch.distributed.is_available() and it returns False

[TimZaman]

Thanks @soumith . I tried that indeed, but I was just wondering if it was documented anywhere, looked all over the place (e.g. maybe another backend would work?). Any tips on using MacOS anyway? It's mostly for convenience, dev testing.
[I am assuming the author of this issue was on MacOS too, so i guess the issue can be closed]
I also noticed TCP has been deprecated, but the docs seem full with it. Am i correct in thinking the docs need updating? I'd be glad to help out.

[soumith]

Do you have multiple GPUs on your OSX machine? :O :O :O

Right now, we dont support CUDA or multi-GPU or a few of the distributed backends (such as gloo) on OSX, but if you are inclined, compile on OSX from source and it should have torch.distributed.deprecated working (i.e. the old distributed backend).

[TimZaman]

I think we can close this issue.

[TimZaman]

Ran into this again on lazy sunday, looks like Pieter added this! pytorch/pytorch#20380

[Undead-Joe]

if i have cuda but i don't want use it.how should i do?