Problem: when sparse=true, _embedding_bag_backward called at::_sparse_coo_tensor_unsafe_symint function which returns a sparse tensor. Since PyTorch/XLA does not support sparse tensors, this resulted in a dispatch error (see the original issue).
Solution: although, ideally we should support sparse tensors, in the short-term we decided (in an offline discussion) to fallback to the dense backwards function.
Fix: #6008
This PR, in addition to https://github.com/pytorch/pytorch/pull/129691, add support for embedding bag calls when its
sparseparameter is true.Problem: when
sparse=true,_embedding_bag_backwardcalledat::_sparse_coo_tensor_unsafe_symintfunction which returns a sparse tensor. Since PyTorch/XLA does not support sparse tensors, this resulted in a dispatch error (see the original issue).Solution: although, ideally we should support sparse tensors, in the short-term we decided (in an offline discussion) to fallback to the dense backwards function.
cc @miladm @JackCaoG