Pedro Cuenca

~~@ajmssc I know! I wanted to wait a few hours in case @tomaarsen wanted to comment, merging now :)~~ lol, actually I can't do it either (or run the workflow),...

We are having the same problem in [🤗 diffusers](https://github.com/huggingface/diffusers/issues/363), but I fail to see where the offending dimensions are being generated or used. The error arises when attempting to run...

In fact, as [@FahimF reports](https://github.com/huggingface/diffusers/issues/363#issuecomment-1237723403), the crash does not occur for larger batches such as those that would result in latent shapes of `(6, 4, 64, 64)` or `(8, 4,...

Regarding the reproducibility code, @patil-suraj replaced `einsum` with `matmul` in the [diffusers codebase](https://github.com/huggingface/diffusers/pull/445/files) and the problem still occurs in exactly the same way. In fact, following up on @Birch-san's example,...

Update after testing PyTorch 1.13.0 from `test` on Ventura 13.0 Beta (22A5373b). - My previous crash repro, and the ones reported by @malfet, now work as reported by @kulinseth :)...

Gently pinging @kulinseth. Do we have confirmation whether this lies on the `mps` side or in PyTorch itself, given that the MPS Graph version worked fine when I checked? https://github.com/pytorch/pytorch/issues/84039#issuecomment-1281125012

Hi @kulinseth, sorry for the ping :) It isn't urgent, but a broad time-frame estimation would be awesome here. Thanks a lot for your work!

Thanks, @kulinseth! Happy to test on any of the nightlies when it makes it there.

@kulinseth My repro above works fine on macOS Ventura 13.3 beta, thanks! However, inference using `diffusers` with some model architectures still fails (but it works for others). For example, the...

Looks good! I think it's very compelling to model the existing xFormers and sliced attention optimizations as just instances of the new "cross-attention processor" class. I also think it would...