[Sana] Add Sana, including SanaPipeline, SanaPAGPipeline, LinearAttentionProcessor, Flow-based DPM-sovler and so on.

[lawrence-cj]

What does this PR do?

This PR will add the official Sana (SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer) into the diffusers lib. Sana first makes the Text-to-Image available on 32x compressed latent space, powered by DC-AE(https://arxiv.org/abs/2410.10733v1) without performance degradation. Also, Sana contains several popular efficiency related techs, like DiT with Linear Attention processor and we use Decoder-only LLM (Gemma-2B-IT) for low GPU requirement and fast speed.

Paper: https://arxiv.org/abs/2410.10629
Original code repo: https://github.com/NVlabs/Sana
Project: https://nvlabs.github.io/Sana

Core contributor of DC-AE:
work with @[email protected]

Core library:

We want to collaborate on this PR together with friends from HF. Feel free to contact me here. Cc: @sayakpaul, @yiyixuxu

Core library:

• Schedulers: @yiyixuxu
• Pipelines and pipeline callbacks: @yiyixuxu and @asomoza
• Docs: @stevhliu and @sayakpaul
• General functionalities: @sayakpaul @yiyixuxu @DN6

HF projects:

• transformers: different repo
• safetensors: different repo

-->

Images is generated by SanaPAGPipeline with FlowDPMSolverMultistepScheduler

[lawrence-cj]

i think in the bf16 repository you have the fp32 weights as well; are those a fp32 copy of the bf16 compatible weights? if so that makes sense but otherwise it may confuse users that don't know to pass variant=bf16 in

Yes. It's just a FP32 copy of BF16 weight and I run it successfully.

[bghira]

without complex human instruction:

with:

is it possible there is something wrong with the CHI implementation here? it makes all images worse.

for example with CHI enabled it's putting 508 tokens of input through the model instead of just 300 (206 from CHI plus the 300 prompt tokens (padded) and i don't know why we need this many tokens. is it supposed to be 300 total?

[lawrence-cj]

What’s your inference code? @bghira

[bghira]

we use encode_prompt via pipeline to save the embed and then pass it back in for inference time so the text encoder can be unloaded first. other than this just using the BF16 weights

[lawrence-cj]

What's your prompt? @bghira

[a-r-r-o-w]

@hlky Would you like to give the changes to schedulers here a review? I'm preparing to merge it shortly after I add the integration tests in the next hour since YiYi has approved and confirmed on Slack. I've tested all the normal models (not the multilingual ones) and they seem to work well (I did the conversions myself when testing, but for the integration tests, I will be using the remote checkpoints and match slices). I have not exhaustively tested all scheduler changes though - only DPMSolverMultistep and FlowMatchEulerDiscrete, but I think that should be okay since it is copied logic (from make fix-copies).

[HuggingFaceDocBuilderDev]

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]

@a-r-r-o-w Scheduler changes look good, thanks

[a-r-r-o-w]

Thank you @lawrence-cj and team! The paper was very insightful and it was very cool to come across the ideas developed.

Thanks for bearing with our reviews too! Will merge the PR once the CI passes

[lawrence-cj]

Thank you so much for your effort! Love you guys. I was stuck by other things, sorry for the late reply! !
@sayakpaul @a-r-r-o-w @bghira @yiyixuxu @hlky