[New model] Support model qwen image layered

[Bounty-hunter]

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

Purpose

Support Qwen image Layered model

Run with image_edit.py，parameters of strong correlation are --color-format "RGBA" and --layers x

Test Plan

(1) run with vllm_omni

python image_edit.py --model "Qwen/Qwen-Image-Layered"   --image xxx --output "image_1_layered_" --num_inference_step 50 --cfg_scale 4.0 --layers x --prompt "" --color-format "RGBA"  --seed 777

(2) run with diffusers

from diffusers import QwenImageLayeredPipeline
import torch
from PIL import Image

pipeline = QwenImageLayeredPipeline.from_pretrained("Qwen/Qwen-Image-Layered")
pipeline = pipeline.to("cuda", torch.bfloat16)
pipeline.set_progress_bar_config(disable=None)

image = Image.open("/home/d00806799/code/vllm-omni/examples/offline_inference/image_to_image/vllm.jpg").convert("RGBA")
inputs = {
    "image": image,
    "generator": torch.Generator(device='cuda').manual_seed(777),
    "true_cfg_scale": 4.0,
    "negative_prompt": " ",
    "num_inference_steps": 50,
    "num_images_per_prompt": 1,
    "layers": 3,
    "resolution": 640,      # Using different bucket (640, 1024) to determine the resolution. For this version, 640 is recommended
    "cfg_normalize": False,  # Whether enable cfg normalization.
    "use_en_prompt": False,  # Automatic caption language if user does not provide caption
}

with torch.inference_mode():
    output = pipeline(**inputs)
    output_image = output.images[0]

for i, image in enumerate(output_image):
    image.save(f"{i}.png")

(3) Run image edit also success

INFO 12-20 14:16:47 [omni_diffusion.py:86] Prepared 1 requests for generation.
INFO 12-20 14:16:47 [diffusion_engine.py:43] Pre-processing completed in 0.0680 seconds
INFO 12-20 14:17:47 [shm_broadcast.py:501] No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g. compilation, weight/kv cache quantization).
INFO 12-20 14:18:43 [diffusion_engine.py:48] Generation completed successfully.
INFO 12-20 14:18:43 [diffusion_engine.py:53] Post-processing completed in 0.0667 seconds
Total generation time: 116.1441 seconds (116144.12 ms)
Saved edited image to /home/d00806799/code/vllm-omni/examples/offline_inference/image_to_image/output_image_edit_0.png
INFO 12-20 14:18:43 [gpu_worker.py:198] Worker 0: Received shutdown message
INFO 12-20 14:18:43 [gpu_worker.py:222] event loop terminated.
INFO 12-20 14:18:43 [gpu_worker.py:253] Worker 0: Shutdown complete.

Test Result

(1) vllm-omni 2-layers

(2) diffusers 2-layers

(3) vllm-omni 3-layers

(4) diffusers 3-layers

---

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)

[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".

[chatgpt-codex-connector]

Use request resolution instead of list attribute

Inside the pre-processing loop the resolution is read from requests rather than the individual req, so the function raises an AttributeError before any request is processed because the list object has no resolution attribute. This prevents the new layered pipeline from calculating dimensions or running at all.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Guidance branch calls time embedding with wrong signature

The guidance code path calls self.time_text_embed(timestep, guidance, hidden_states, additional_t_cond), but QwenTimestepProjEmbeddings.forward only accepts (timestep, hidden_states, addition_t_cond=None). When guidance is enabled (e.g., for guidance-distilled models), this path raises a TypeError before any diffusion steps run.

Useful? React with 👍 / 👎.

[hsliuustc0106]

add test ci please

[ZJY0516]

please also test if this break qwen-image and qwen-image-edit

[ZJY0516]

We can import this from diffusers directly

[Bounty-hunter]

The modify for image layered in autoencoder_kl_qwenimage.py only in main branch; and current diffusers release version is 0.36

[ZJY0516]

User should install latest diffusers from source as mentioned in https://huggingface.co/Qwen/Qwen-Image-Layered#quick-start

[ZJY0516]

remove this file please

[hsliuustc0106]

shall we leave it for a later PR fixing this?

[hsliuustc0106]

add the test result and provide the run example command

[Bounty-hunter]

add the test result and provide the run example command
done

[hsliuustc0106]

lgtm