[Manual Test] diffusion endpoints allow model mismatch (upstream #2805)

[lishunyang12]

Manual test of the 3-tier model routing.

How to use

Apply one of these labels to kick off a review:

Label	Model	When
claude-review	Opus 4.7	Deep review
claude-quick	Haiku 4.5	Fast LGTM-style triage

Then in comments, talk to the reviewer:

• @claude review — full review (Sonnet)
• @claude why did you flag X? — question (Sonnet)
• @claude I disagree — Y handles that — disagreement (Sonnet)
• @claude LGTM? — short verdict (Sonnet)
• @claude look again after my fix — re-review (Sonnet)

Good comparison experiment

1. Apply claude-quick first → Haiku does a fast scan
2. Then apply claude-review → Opus does a deep pass
3. Compare the two results

Mirrors upstream vllm-project#2805 ([Bugfix] diffusion endpoints allow model mismatch — 3 files / 61 lines). Do not merge.

[lishunyang12]

@claude review

[claude]

LGTM

[claude]

The or request.model fallback silently defeats the new hard-error guard. When both handler.model_name and app_model_name are None (which is possible if OmniOpenAIServingVideo was constructed via __init__ with the default model_name=None, e.g. not via for_diffusion), effective_model_name becomes request.model and the check request.model != effective_model_name is always False — no 400 is ever raised.

The image endpoints don't have this hole: _get_engine_and_model uses "unknown" as the fallback, not the client's model string.

If the intent is a hard reject on mismatch, pull the model name directly from the handler/app and error explicitly when it's unset, rather than falling back to what the client sent:

Suggested change

	effective_model_name = handler.model_name or app_model_name or request.model or "unknown"
	effective_model_name = handler.model_name or app_model_name
	if effective_model_name is None:
	raise HTTPException(
	status_code=HTTPStatus.SERVICE_UNAVAILABLE.value,
	detail="Server model name is not configured.",
	)

Then keep the existing mismatch check below unchanged.

[claude]

The or request.model or "unknown" fallback defeats the validation. When both handler.model_name and app_model_name are None, effective_model_name falls through to request.model, so request.model != effective_model_name is always False — the 400 below can never fire. The image endpoints don't have this hole because they compare directly against model_name from _get_engine_and_model.

Suggested change

	effective_model_name = handler.model_name or app_model_name or request.model or "unknown"
	if request.model is not None and effective_model_name is not None and request.model != effective_model_name:
	effective_model_name = handler.model_name or app_model_name
	if request.model is not None and effective_model_name is not None and request.model != effective_model_name:

[lishunyang12]

Closing — superseded by label-only config change (no @claude, 3 tiers).

[claude]

👍