torchrec
torchrec copied to clipboard
remember split_embedding_weights
Summary: ATT currently calling split_embedding_weights repeatedly over FBGEMM kernels provide tensor views that are identical, but have different ids (hash values). As a result, it makes comparing tensors from state_dict tricky
Differential Revision: D37337616
This pull request was exported from Phabricator. Differential Revision: D37337616