Brendan Folie
Brendan Folie
All collectives have been implemented in eager mode. Not all work in compiled mode but we have the same coverage as PT native does.
> XLA doesn't have an explicit control on resharding a tensor, but this can be achieved by simply cloning the existing tensor and do a `mark_sharding` on new tensor Is...
Thanks @rpsilva-aws, yes that is the discussion I was thinking of
Thanks for this proposal and context @fhaolinaws. As we've discussed I think this is a good approach. It improves the user experience, provides both collective-based and XLA-based options, and should...
The approach seems good. However I don't think this should be merged until the corresponding PyTorch PR has been merged and these tests can be activated.