Sasha Rush

It should just work? Is there a bug?

I see what you mean. So you are suggesting writing another intermediate operator that directly does both combinations without storing intermediates. Let's do this together. Maybe you can give a...

Or perhaps you are just suggesting that genbmm should support broadcasting along the first dimension? Would that work? (A->BC could be size (1, NT, NT * NT) and still be...

I like the second solution better. If you are motivated to give it a try, here's how to do it. 1) edit this line so you check both the size...

Cool, yeah I played with keops a bit but it didn't performs as well as I would have liked. (see https://github.com/harvardnlp/pytorch-struct/blob/master/torch_struct/semirings/keops.py ) . But I think maybe that was because...

Btw, does this same issue appear for dependency parsing? It would be nice to have a kernel that wasn't so CFG specific.

Oh no, you should definitely not try to do triplets in CUDA that would be really messy. I think the right way to do this is to remove this expansion...

Which algorithm are you talking about particularly? Also what do you mean by linear scan here? I don't use linear-scan for any of the tree approaches. I started by implementing...

Hmm, would be curious to know how CKY_CRF with logbmm compares to manual backward. Not sure where it is storing so much extra memory.

One really nice trick to save memory (without more code) is by recomputing is to use Checkpointing. It basically just automatically reruns forward for you. Here is an example of...