fix import container_abcs issue

[ShoufaChen]

Nightly PyTorch has removed container_abcs from torch._six.
pytorch/pytorch@58eb233#diff-b3c160475f0fbe8ad50310f92d3534172ba98203387a962b7dc8f4a23b15cf4dL35

Also fix #1048 and #1046

[ShoufaChen]

CC @mcarilli @seryilmaz @definitelynotmcarilli

Hi, would you mind having a review?

[mcarilli]

Apex is in maintenance mode at best. I'm not working on apex, but i guess this counts as "maintenance." I AM working to ensure all of apex's commonly used features have better support through Pytorch itself. I strongly recommend anyone still using apex amp or DDP to use torch.cuda.amp and torch.nn.parallel.DistributedDataParallel instead. torch.optim fused optimizers are also coming soon courtesy of @izdeby.

[mcarilli]

I'd rather keep existing import for Pytorch 1.7 and earlier:

TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])

if TORCH_MAJOR == 1 and TORCH_MINOR < 8:
    from torch._six import container_abcs
else:
    import collections.abc as container_abcs

[ShoufaChen]

Thanks for your suggestion. I've updated it.

[dbonner]

Thanks @ShoufaChen and @mcarilli for such a prompt response.
I can confirm that your patch fixes the import isssue on my machines.
All the best,
Dan