[Diffusion][Feature] CFG parallel support for Qwen-Image

[wtomin]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

This PR aims to support CFG-Parallel for Qwen-Image series of models. CFG-parallel runs the positive/negative prompts of classifier-free guidance (CFG) on different devices, then merges on a single device to perform the scheduler step. It includes:

• Edits the pipelines: split the positive/negative prompts in diffuse function, and merge the predicted noise before scheduler step.
  pipeline_qwen_image.py
  pipeline_qwen_image_edit.py
  pipeline_qwen_image_edit_plus.py
  pipeline_qwen_image_layered.py

• Edits the image_edit.py and text_to_image.py to support cfg_parallel_size argument.

• Update documents.

Test Plan

• image generation

python examples/offline_inference/text_to_image/text_to_image.py --cfg_parallel_size 2

python examples/offline_inference/text_to_image/text_to_image.py --cfg_parallel_size 2 --cache_backend tea_cache

• image edit

python examples/offline_inference/image_to_image/image_edit.py --model "Qwen/Qwen-Image-Edit" --image ./qwen_image_output.png --prompt "turn this coffee cup to a glass of wine'" --output output_image_edit.png --num_inference_steps 50 --cfg_scale 4.0 --cfg_parallel_size 2

python examples/offline_inference/image_to_image/image_edit.py --model "Qwen/Qwen-Image-Edit" --image ./qwen_image_output.png --prompt "turn this coffee cup to a glass of wine'" --output output_image_edit.png --num_inference_steps 50 --cfg_scale 4.0 --cache_backend tea_cache --cfg_parallel_size 2

python examples/offline_inference/image_to_image/image_edit.py --model "Qwen/Qwen-Image-Edit-2509" --image ./qwen_image_output.png --prompt "turn this coffee cup to a glass of wine'" --output output_image_edit.png --num_inference_steps 50 --cfg_scale 4.0 --cfg_parallel_size 2

python examples/offline_inference/image_to_image/image_edit.py --model "Qwen/Qwen-Image-Layered" --image ./qwen_image_output.png --prompt "turn this coffee cup to a glass of wine'" --output output_image_edit.png --num_inference_steps 50 --cfg_scale 4.0 --layers 2 --color-format "RGBA" --output "layered" --cfg_parallel_size 2

Test Result

task	model	cfg_parallel_size	time	generated image
T2I	Qwen/Qwen-Image	1	20.5s
T2I	Qwen/Qwen-Image	2	13.08s
I2I	Qwen/Qwen-Image-Edit	1	54.3s
I2I	Qwen/Qwen-Image-Edit	2	29.9s
I2I	Qwen/Qwen-Image-Edit-2509	1	45.0s
I2I	Qwen/Qwen-Image-Edit-2509	2	25.5s
I2I	Qwen/Qwen-Image-Layered	1	32.3s
I2I	Qwen/Qwen-Image-Layered	2	19.3s

task	cache backend	model	cfg_parallel_size	time	generated image
T2I	tea_cache	Qwen/Qwen-Image	1	12.5s
T2I	tea_cache	Qwen/Qwen-Image	2	10.6s
I2I	tea_cache	Qwen/Qwen-Image-Edit	1	24.35s
I2I	tea_cache	Qwen/Qwen-Image-Edit	2	17.06s

Setting:

• vllm: 0.12.0
• pytorch: 2.9.0
• python: 3.12
• cuda: 12.8

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[ZJY0516]

looking forward to it

[hsliuustc0106]

do we expect it to be merged before 1230 release? @wtomin

[wtomin]

do we expect it to be merged before 1230 release? @wtomin

I think so. I will get it done today.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[hsliuustc0106]

any progress or comment @gcanlin @ZJY0516 @wtomin @SamitHuang

[gcanlin]

Would CGF parallel perform better than USP? If roughly the same, I’d prefer to move this to the next release to ensure the quality of 0.12.0.

[ZJY0516]

Would CGF parallel perform better than USP? If roughly the same, I’d prefer to move this to the next release to ensure the quality of 0.12.0.

They are orthogonal

[wtomin]

@hsliuustc0106 The current branch is compatible with tea_cache.

[hsliuustc0106]

fix precommit please

[hsliuustc0106]

@ZJY0516 PTAL final check

[gcanlin]

Also work on NPU. Thanks!

python text_to_image.py --cfg_parallel_size 2

And it achieved the expected speedup.

model	config	time
Qwen-Image	--cfg_parallel_size 2	58s
Qwen-Image	--cfg_parallel_size 1	37s

[hsliuustc0106]

lgtm

[david6666666]

@wtomin lack of parameter passing in online serving scenarios, please add