it was recommended to create a new AttentionProcessor for ptxla. Unfortunately that one doesn't work for Flux and I feel its a lot of duplication of code if I do the same for Flux as it would almost be a copy/paste of FluxAttentionProcessors. Please advice.

Okay for me to create a new class even if that means duplication. Once we start to more formally support PyTorch XLA (i.e., through inclusion in our CI), we can start revisiting them to refactor but for now, new classes are completely fine by me. Cc: @yiyixuxu WDYT?

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

@hlky @sayakpaul created new class and addressed the comments + ran make style/quality. Please take another look.

@entrpn could you run make style && make quality?

make style && make quality

done

Thanks for contributing this!

Sorry if I've misunderstood, as I don't understand this very well, but was this change in FluxAttnProcessor2_0 accidental?

https://github.com/huggingface/diffusers/pull/10515/files#diff-96aeb208ac8e69bdb368e0a481a0cd50173b3f4bf19e5da500c251614820c8cdR2327

I think it reverts this change?

• pass attn mask arg for flux #10122

@josephrocca yes this should not have changed. I'll need to make a PR to revert this change.