Bhavya Bahl
Bhavya Bahl
> Is this the only operation missing support for `long` `dtype`? I assume this is TPU specific? On more testing, I found when one of the tensors is a long...
@wonjoolee95 This should be ready for review. I implemented the _embedding_bag_forward_only op for now. It seems that a lot of tests fail without the backward pass implementation.
@wonjoolee95 Do you think that you have time to look at this PR?
Looking at this now
I'll wait for the new nightly release and confirm that this issue is fixed before closing.
#53, #67 adds the interface for tokenizer. Still need to work on ResultTokens.
PR: https://github.com/pytorch/xla/pull/8202 Reason: Fix dataloading issue for SD2 model Cherry-pick: https://github.com/pytorch/xla/pull/8212
@johnsutor This error happens usually when there's a mismatch between torch and torch_xla versions installed. Maybe also update the pytorch to nightly version. ``` pip3 install --pre torch torchvision torchaudio...
Thanks for the change! I am going to run the CI and see if it breaks anything
Can you also please add a test for embedding backward similar to this one https://github.com/pytorch/xla/blob/193f7049ef56639122b11c03add44512aee9f66a/test/test_operations.py#L738