Closed tugsbayasgalan closed 1 year ago
cc: @bdhirsh @Chillee @gmagogsfm @larryliu0820
This is a composite compliance issue. It looks like we actually have a composite-compliant decomposition for ctc_loss
, but it currently only runs for tensor subclasses, and doesn't run when functionalization is active (which I think we should fix). More discussion here: https://github.com/pytorch/pytorch/pull/84752/files#r989417906
🐛 Describe the bug
To repro:
which gives following error:
I think this is because ctc_loss is still not composite_complaint even after this diff (https://github.com/pytorch/pytorch/pull/84752/files).
Versions
latest master
cc @bdhirsh @ezyang @soumith @SherlockNoMad @ngimel