Add Qwen-Image

[atakaha]

Enable qwen-image with Gaudi specific optimization. Rewrite RoPE from complex to cos/sin.

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[regisss]

Nice!

Can you add a new entry for this model in the tables here and there please?

Can you also add some tests? Those can be directly taken and adapted from Diffusers.

Ccing @dsocek too

[dsocek]

@atakaha looks quite good to me, I agree with @regisss please update high level docs too add CI tests from diffusers and possibly 1 slow test (for inference regressions check)

[atakaha]

observing regression after rebase. Need to fix the regression.

[atakaha]

@regisss and @dsocek please review this PR

• Update required files.
• Add slow pytest and confirmed it pass
• rebased and fixed conflicts
• confirmed both lazy and eager mode outputs.

[dsocek]

Looks good, small fix suggested to add HPU centric example doc string

[regisss]

Could we do this monkey patching directly in the __init__ of the GaudiDiffusionPipeline?

optimum-habana/optimum/habana/diffusers/pipelines/pipeline_utils.py

Line 145 in 840e497

# Patch for unconditional image generation

[atakaha]

I added like below, but didn't work.
from ..models.transformers import ( GaudiQwenTimestepProjEmbeddings, GaudiQwenEmbedRope, GaudiQwenDoubleStreamAttnProcessor2_0 ) diffusers.models.transformers.transformer_qwenimage.QwenTimestepProjEmbeddings = GaudiQwenTimestepProjEmbeddings diffusers.models.transformers.transformer_qwenimage.QwenEmbedRope = GaudiQwenEmbedRope diffusers.models.transformers.transformer_qwenimage.QwenDoubleStreamAttnProcessor2_0 = GaudiQwenDoubleStreamAttnProcessor2_0

[atakaha]

monkey patch doesn't work in __init__ but from_pretrained() of the GaudiDiffusionPipeline. Placed it in from_pretrained().

[regisss]

Sounds good, thanks for looking further into it!

[github-actions]

The code quality check failed, please run make style.

[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.

[regisss]

I left one more comment to basically move all the logic in modeling_utils.py to the from_pretrained method.

Also, the code style check failed, please run

make style

from the root of the repo.

[regisss]

LGTM!