Richard Zou

https://github.com/pytorch/functorch/pull/901 deletes the cross_entropy monkey patch. The other monkeypatches need some work in PyTorch core

Merged

This may be a bit difficult to do in a performant manner. One thing we could do is decompose the GRU and rely on AOTAutograd to generate a fast kernel...

potentially related: https://github.com/pytorch/pytorch/issues/82577, cc @samdow

I'm not sure what is going on here, but we'll take a deeper look

I'm not sure -- does _forward_pre_hooks get copied by deepcopy? In general functorch doesn't work very well with module hooks at the moment

Thanks for the issue, @teddykoker. This seems like a reasonable API. We've had some decision paralysis in the past on what the best API for this would look like, but...

We don't have to support det for now for that test, so please feel free to merge https://github.com/pytorch/pytorch/pull/80217 and we can worry about that test later.

@Lezcano will https://github.com/pytorch/pytorch/pull/80217 land soon? If not, since functorch is in pytorch/pytorch now, I want to reenable the above tests to prevent regressions

Cool, going to re-enable them. If it's not too difficult we should try to fix them in the PR to avoid regressions