[feature] Add context parallel cp_plan support to many mainstream models

[DefTruth]

I am working in cache-dit, where I use diffusers as the native context parallel backend and extensively utilize custom cp_plan to support context parallel for mainstream models (and has also make many models worked with context parallel + attn_mask via native attn backend). These models include: FLUX, Qwen-Image, LTXVideo, Wan, HunyuanImage, HunyuanVideo, CogVideoX, CogView3Plus, CogView4, Qwen-Image-Lightning, ConsisID, Chroma, VisualCloze, etc. If you wish to integrate these implementations into diffusers, I am willing to submit some PRs to provide support. I will keep a long-term focus on the context parallel feature of diffusers and hope it can become an extremely user-friendly function. our experimental cp_planners are at: cp_planners

@sayakpaul @yiyixuxu @DN6

[sayakpaul]

Thanks for keeping us in your thoughts!

Yes, I think our custom cp_plan argument is perfect for those use cases and more sustainable in terms of maintenance. I think on our end, we will likely start adding lightweight test suites to check the effectiveness of passing custom cp_plan when enabling parallelism.

Also note that the default cp_plan is enabled for ultra-popular models, and that list is quite smaller than what you mentioned. So, we will continue to ensure custom cp_plan works the way it should work and you could play a crucial part in helping us test that.

[DefTruth]

Get~ I will provide some PRs later to submit the custom cp_plan implemented in cache-dit to diffusers to support more models in using context parallelism (like HunyuanImage, HunyuanVideo, CogVideoX, CogView3Plus, CogView4, etc). You can determine whether to merge it based on the actual situation.