[BugFix]config priority fix

[Bounty-hunter]

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

Background

Currently, vllm-omni constructs stage configurations during startup according to the following priority rules.

Diffusion Model

• If stage_configs_path is provided, construct the configuration from it.
• Otherwise, construct the configuration from CLI kwargs, e.g.:

python text_to_image.py --tensor-parallel-size 4

Omni Model

• If stage_configs_path is provided, construct the configuration from it.
• Otherwise, load the configuration from the default locations:
  vllm_omni/model_executor/stage_configs/
vllm_omni/platforms/xxx/stage_configs/

Problem

HunyuanImage-3.0 supports two execution modes:

• DIT only
• AR + DIT (multi-stage)

However, during startup the system always loads the stage configuration from
vllm_omni/model_executor/stage_configs/. As a result, it always initializes the AR stage (or AR + DIT in the future), even when only the DIT stage is required.

PR #1826
attempts to address this by placing both AR and DIT configurations in hunyuan_image_3_moe.yaml, and dynamically selecting the relevant configuration based on the task type (text-to-image, image-to-text, etc.).

However, this approach introduces a new issue:

When starting Hunyuan DIT, a YAML configuration must always be specified, or use the default config with 8 card and 8 tensor parallel from hunyuan_image_3_moe.yaml

CLI kwargs (e.g. --tensor-parallel-size) are totally ignored.

This problem is discussed in:
#2282

Purpose

For DIT-only models (e.g., HunyuanImage DIT), the startup behavior should be consistent with other diffusion models. Users should be able to launch the model directly with CLI arguments, for example:

python -u examples/offline_inference/text_to_image/text_to_image.py \
--model /home/models/tencent/HunyuanImage-3.0  \
--prompt "A brown and white dog is running on the grass"  \
--output output_image_latest.png    \
--num-inference-steps 50    \
--tensor-parallel-size 4    \
--diffusion-only \
--cfg-scale 4.0

For multi-stage models (e.g., HunyuanImage AR + DIT), the system should still load stage configurations from the default directories:vllm_omni/model_executor/stage_configs/ or vllm_omni/platforms/xxx/stage_configs/.

Change

Introduce a new parameter: --diffusion-only

• If --diffusion-only is set: it default to construct config from kwargs
• If --diffusion-only is not set: it default to construct from vllm_omni/model_executor/stage_configs/ or vllm_omni/platforms/xxx/stage_configs/

cc @fake0fan @xuechendi @yinpeiqi

Test Plan

Test Result

---

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: d03ed7022f

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]

Keep diffusion stage in default Hunyuan config

This change removes the stage_id: 1 diffusion stage from the default hunyuan_image_3_moe config, so launching without diffusion_only now resolves to an AR-only pipeline and breaks image generation flows that previously relied on default config resolution. That regresses both offline usage (tests/e2e/offline_inference/test_expert_parallel.py expects Omni(model="tencent/HunyuanImage-3.0") to return images) and serving paths (vllm_omni/entrypoints/openai/api_server.py rejects pipelines with no diffusion stage for /v1/images/*). The diffusion stage should remain in the default config, with diffusion_only controlling config-construction priority rather than deleting the stage.

Useful? React with 👍 / 👎.

[Bounty-hunter]

fix by #2076

[Bounty-hunter]

fix by #2076