`WanVACETransformer3DModel` does not inherit `AttentionMixin` while `WanTransformer3DModel` does

[J4BEZ]

Hello,

While installing the latest version of diffusers from source and implementing a custom Wan VACE pipeline, I noticed a potential inconsistency that I hope may be helpful to point out.

WanTransformer3DModel inherits the following:
ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin, and AttentionMixin. link

However, WanVACETransformer3DModel inherits all of the above except for AttentionMixin. link

This difference does not cause major issues for video generation or LoRA application, but when attempting to call the transformer.set_attn_processor() method in a custom pipeline, I encountered an error with WanVACETransformer3DModel, unlike WanTransformer3DModel.

Thank you very much for all the incredible work and dedication you put into maintaining the diffusers library, which makes it much easier for researchers and developers to enjoy working with generative models. Wishing you good health and a peaceful day ahead!

Regards, J4BEZ

[DN6]

Hi @J4BEZ yes, we do need to refactor VACE to use the AttentionMixin. Would it be possible for you to open a PR for it?

[J4BEZ]

@DN6 Yes, I would be glad to contribute to diffusers again.
Thank you very much for your kind reply😄

[DN6]

Thanks @J4BEZ. Closing since #12268 was merged.

[J4BEZ]

@DN6 Thank you! I hope you a peaceful day 😄