[Manual Test] Bagel multi-stage cfg bugfix (upstream #2801)

[lishunyang12]

Test the maintainer-style review bot with 3 tiers.

Mirrors upstream vllm-project#2801 — [BugFix]: Fix multi-stage cfg bug (4 files / 106 lines, Bagel diffusion pipeline + stage input processor + 2 tests).

Try all three tiers

Apply each label in turn (wait for each to finish before adding the next):

Order	Label	What to expect
1	claude-quick	Haiku posts a single top-level preview: what-this-does summary, meta check, file-by-file one-liners, depth recommendation. No inline bugs.
2	claude-review	Sonnet does a full maintainer review — Phase 0 meta audit, Phase 1 bugfix-category checks, Phase 2 code review, Phase 3 test audit. As many inline comments as it finds warranted.
3	claude-deep	Opus does the same job as claude-review, but with stronger reasoning for subtle cross-file concerns.

What to compare

• Does claude-quick actually help you decide whether to do a deep review? Or is it fluff?
• Does claude-review match what a human maintainer would flag?
• Does claude-deep find things claude-review missed?

Do not merge.

[claude]

Preview (Haiku)

What this PR does: Fixes a multi-stage CFG (Classifier-Free Guidance) bug in the Bagel diffusion pipeline where the condition for injecting KV cache was incorrect, and updates test reference pixel values to match the corrected behavior.

Meta check:

• Title prefix: [Manual Test] — non-standard (should be [Bugfix], [Model], etc.). Note: PR is explicitly a test harness for the review bot; the actual upstream fix is labeled [BugFix].
• PR description: Filled — explains this mirrors upstream [BugFix]: Fix multi-stage cfg bug vllm-project/vllm-omni#2801 and tests the review bot tiers.
• DCO sign-off: Present on commit 1e69e9f.
• Size: 54 additions, 52 deletions (~106 LOC) — under 500, no RFC required.

Files touched:

• tests/e2e/offline_inference/test_bagel_img2img.py — Updates REFERENCE_PIXELS RGB values for CUDA and ROCm test paths.
• tests/e2e/offline_inference/test_bagel_text2img.py — Updates REFERENCE_PIXELS RGB values for CUDA and ROCm test paths.
• vllm_omni/diffusion/models/bagel/pipeline_bagel.py — Fixes CFG branch logic (changes condition from cfg_img_kv is None and cfg_text_kv is not None to cfg_img_kv is None), adds explanatory comment for text2img empty-KV path, removes CFG-disabled fallback warning.
• vllm_omni/model_executor/stage_input_processors/bagel.py — Adds early returns when neg_prompt is empty in expand_cfg_prompts and expand_cfg_prompts_think; changes default empty negative prompt from "<|im_start|><|im_end|>" to "".

Depth recommendation: claude-review — This is a [Bugfix] for a Bagel pipeline CFG path with test updates. Needs full maintainer review to verify the logic fix addresses root cause (not symptom), that test updates reflect correct output, and that regression tests are adequate.

[lishunyang12]

Closing — preview prompt tightened. New test PR to follow.