[Quantization] Enable FP8 online quantization for Qwen-image-edit text encoder

[xRay2016]

ref #1854

Purpose

Support FP8 quantization for the Qwen-Image-Edit text encoder part by following #1338.

This PR updates the Qwen Image Edit diffusion pipeline to:

• Load text_encoder, transformer, and vae through component weight sources.
• Instantiate the text encoder with create_transformers_model(...) so vLLM-Omni quantization replacement can be applied.
• Add a root_prefix path for recursive linear replacement, allowing quantization config entries such as text_encoder.
• Map HuggingFace Qwen-Image-Edit text encoder weight names to the vLLM model structure during weight loading.
• Load the VAE from config so it is populated through the unified weight loader.

It also adds a quantization quality test case for Qwen/Qwen-Image-Edit with FP8 applied to the text encoder.

Test Plan

Run the targeted full-model quality test:

VLLM_OMNI_QUALITY_OUTPUT_DIR=/tmp/modelopt_quality_outputs \
.venv/bin/python -m pytest tests/diffusion/quantization/test_quantization_quality.py \
  -v -m "" -k "fp8_qwen_image_edit_text_encoder"

Run the e2e qwen image edit test:

.venv/bin/python -m pytest tests/e2e/accuracy/test_qwen_image_edit.py::test_qwen_image_edit_single_matches_diffusers -v -s

Test Result

The result of full-model quality test is:

============================================================
Quantization Quality: fp8_qwen_image_edit_text_encoder
============================================================
  Baseline:      Qwen/Qwen-Image-Edit
  Quantized:     Qwen/Qwen-Image-Edit
  Method:        {'text_encoder': {'method': 'fp8'}, 'transformer': None, 'vae': None}
  LPIPS:         0.0167  (threshold: 0.15)
  PSNR:          33.3159 dB  (higher is better)
  MAE:           0.014689  (lower is better)
  BF16 memory:   56.87 GiB
  Quant memory:  49.58 GiB  (13% reduction)
  Result:        PASS
============================================================

PASSED

PASSED

The result of e2e qwen image edit is:

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
--- Running Summary
============== 1 passed, 19 warnings in 188.01s (0:03:08) ==============

---

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. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (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.

Reviewed commit: b2b8d2fe73

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]

LGTM. This enables per-component quantization for diffusion models, which is a valuable capability.

The parameter approach is clean and enables quantization config entries like to work correctly. Weight mapping from HF keys to vLLM structure is also well done.

[hsliuustc0106]

@fhfuih PTAL

[xRay2016]

@hsliuustc0106 @fhfuih Hi, do you have any remaining comments on this PR? If not, could you please approve it so we can merge? Thanks!

[hsliuustc0106]

@david6666666 @lishunyang12 PTAL

[lishunyang12]

Show visual comparision

[xRay2016]

Show visual comparision

Hi, @lishunyang12

Here are visual comparison generated by tests/diffusion/quantization/test_quantization_quality.py.

Both images use the same input image, prompt, seed, resolution, inference steps, and sampling settings. I don’t observe obvious visual degradation from the quantized output in these examples.

BF16 baseline

FP8 text encoder

[fhfuih]

I am not quite familiar with quantization. But referring #1338 this PR looks generally fine. Left some small comments on tests

[xRay2016]

Hi! Just following up on this PR. I’ve addressed the previous feedback and would appreciate another review when you have time.

Please let me know if there’s anything else you’d like me to change. If everything looks good, an approval would be appreciated. Thanks!