Hi, thanks for releasing the code. Why do you choose `MPI adaptor` instead of using default `torch.distributed` (just like in [Mask2Former](https://github.com/facebookresearch/Mask2Former/blob/9b0651c6c1d5b3af2e6da0589b719c514ec0d69a/train_net.py#L321C8-L321C8)) ? ```[tasklist] ### Tasks ```
Hi, thanks for releasing the code.
Why do you choose
MPI adaptorinstead of using defaulttorch.distributed(just like in Mask2Former) ?