[Bugfix] Limit Qwen-Image-Edit-2511 input image count

[david6666666]

Summary

• limit QwenImageEditPlusPipeline to at most 4 input images during pre-processing
• fail early with a clear validation error instead of reaching deeper OOM / sequence-length failures
• add a unit test covering the over-limit case

Root Cause

Qwen-Image-Edit-2511 accepts multi-image inputs, but very large image counts can blow past the practical prompt/conditioning limits for this pipeline and eventually surface as OOM or deeper runtime failures. The fastest and smallest safe fix is to reject oversized requests at the input validation boundary.

Why This Fix

This keeps the change minimal and low-risk:

• one early validation gate in the existing pre-processing path
• no changes to the inference core
• clearer user-facing failure mode

Validation

• pytest -q tests/diffusion/models/qwen_image/test_qwen_image_edit_plus.py

Test Plan

• Start a local API server with the local Qwen-Image-Edit-2511 checkpoint via:
  CUDA_VISIBLE_DEVICES=7 PYTHONPATH=/mnt/data4/cwq/worktree/issue2793-qwen-image-edit-oom python -m vllm_omni.entrypoints.cli.main serve /mnt/data1/huggingface/hub/models--Qwen--Qwen-Image-Edit-2511/snapshots/6f3ccc0b56e431dc6a0c2b2039706d7d26f22cb9 --omni --port 8023 --uvicorn-log-level warning
• Send a /v1/images/edits request with 5 input images via curl and verify the request is rejected with a 400 validation error.
• Send a /v1/images/edits request with 4 input images via curl and verify the request succeeds and returns a generated image payload.

Test Result

• pytest -q tests/diffusion/models/qwen_image/test_qwen_image_edit_plus.py
  • Passed.
• E2E with vllm serve + curl against local Qwen-Image-Edit-2511 on GPU 7:
  • 5-image request returned 400 with:
    Received 5 input images. At most 4 images are supported by this model.
  • 4-image request returned 200 and produced a valid 512x512 PNG response.

Fixes #2793.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[hsliuustc0106]

Blocking Issues

1. [Reliability/Safety] vllm_omni/entrypoints/openai/api_server.py:1669 - _get_max_edit_input_images hardcodes return 4 without any model-specific lookup. This function should query the OD config or diffusion pipeline for the actual limit per model, otherwise future models with different limits will break or need manual updates to this helper.

---

VERDICT: REQUEST_CHANGES

The validation logic is correct, but _get_max_edit_input_images hardcodes return 4. This should be model-configurable - either query the OD config or the diffusion pipeline's limit instead of hardcoding per PR.

Why is 4 the right limit? The PR mentions "practical prompt/conditioning limits" but doesn't show calculations. Consider adding a comment or linking to the issue discussion explaining the sequence-length/math behind this threshold.

[david6666666]

Blocking Issues

1. [Reliability/Safety] vllm_omni/entrypoints/openai/api_server.py:1669 - _get_max_edit_input_images hardcodes return 4 without any model-specific lookup. This function should query the OD config or diffusion pipeline for the actual limit per model, otherwise future models with different limits will break or need manual updates to this helper.

VERDICT: REQUEST_CHANGES

The validation logic is correct, but _get_max_edit_input_images hardcodes return 4. This should be model-configurable - either query the OD config or the diffusion pipeline's limit instead of hardcoding per PR.

Why is 4 the right limit? The PR mentions "practical prompt/conditioning limits" but doesn't show calculations. Consider adding a comment or linking to the issue discussion explaining the sequence-length/math behind this threshold.

fixed

[lishunyang12]

Looks good overall -- clean, minimal, and well-tested. The dual-layer validation (API server + pipeline) is the right approach. A few observations:

Positive

• Early rejection before _load_input_images is a smart optimization -- avoids decoding/fetching images that will be rejected anyway.
• Extracting _get_diffusion_od_config as a shared helper is a nice refactor that removes duplication.
• Good test coverage: unit test on the pipeline pre-process path, plus two API-level tests (one confirming _load_input_images is never called).

Minor suggestions (non-blocking)

1. _get_max_edit_input_images string matching is fragile. The "Qwen-Image-Edit-2511" in identifier substring check will match unrelated model names that happen to contain that substring (unlikely today, but brittle). Consider comparing against the canonical HF repo ID or using identifier.endswith(...) / an exact match against a known set.

2. _get_diffusion_od_config is called twice when images exceed the limit: once inside _supports_multimodal_image_inputs and once directly in _get_max_edit_input_images. This is fine for correctness (cheap call), but you could cache the result in a local variable if you want to be tidy.

3. The pipeline-level ValueError vs. the API-level HTTPException. If someone bypasses the API server and calls the pipeline directly with 5 images, they get a ValueError. That's reasonable, but worth noting that the two error messages are slightly different in wording ("Received 5 input images. At most 4..." vs "Received 5 input images. At most 4...") -- actually they match, which is good. Just confirming they stay in sync since the constant is shared.

4. Test file test_qwen_image_edit_plus.py -- mock VAE config is minimal. The test writes {"z_dim": 16} which is enough today, but if get_qwen_image_edit_plus_pre_process_func ever reads additional VAE config keys at init time, this test will break with a confusing error. A small comment in the test noting this is intentionally minimal would help future maintainers.

LGTM -- approving.

[SamitHuang]

The dual-layer validation prevents OOM effectively. However, the model name is still hardcoded in api_server.py, which makes it brittle for future models and violates separation of concerns. Please fix this architectural issue before merging.

[SamitHuang]

This string matching is fragile and hardcodes model-specific logic in the API server. Consider adding a generic attribute like max_multimodal_image_inputs to OmniDiffusionConfig or the model's configuration. This will keep the API server model-agnostic and prevent manual updates for future pipelines.

[SamitHuang]

_get_diffusion_od_config is called twice, once here and once inside _supports_multimodal_image_inputs. Fetch od_config once at the beginning of the function to avoid redundant calls. You can check getattr(od_config, 'supports_multimodal_inputs', False) directly.

[SamitHuang]

This mock VAE config is extremely minimal. Add a brief comment indicating it is intentionally minimal. This helps future maintainers understand why the test might break if get_qwen_image_edit_plus_pre_process_func starts reading more keys at initialization.

[david6666666]

The dual-layer validation prevents OOM effectively. However, the model name is still hardcoded in api_server.py, which makes it brittle for future models and violates separation of concerns. Please fix this architectural issue before merging.

fixed

[SamitHuang]

LGTM, pls fix the CI error

[Gaohan123]

Please fix CI failure. Thanks