pytorch
pytorch copied to clipboard
Mitigates SWDEV-459618
_c10d_functional_autograd::all_to_all_single
seems not implemented on ROCm.
Note: another unfixed problem is the mismatching of outputs between torch.ops.aten._scaled_dot_product_flash_attention and _scaled_dot_product_chunk_flash_attention. We need fix both problems to enable this UT.
Fixes SWDEV-459618