[Bugfix][Qwen-Image] Restore RotaryEmbedding CUDA RoPE for Diffusers e2e

[NumberWan]

Summary

• Fixes [Bug]: Nightly CI, tests/e2e/accuracy/test_qwen_image.py::test_qwen_image_matches_diffusers, AssertionError: PSNR below threshold for Qwen/Qwen-Image: got 26.130234, expected >= 27.000000. #7494: nightly test_qwen_image_matches_diffusers PSNR 26.13 vs gate 27.
• [Rebase] Rebase to vLLM 0.29.0 #7230 switched CUDA Qwen-Image RoPE to _apply_qwen_image_rotary_emb (FP32 complex multiply) inside regional compile of QwenImageTransformerBlock. The helper matches Diffusers apply_rotary_emb_qwen(..., use_real=False) (unit test atol=0), but this e2e test scores Omni vs Diffusers pipeline images, not that helper. Inductor also cannot codegen complex ops.
• Forward uses the pre-[Rebase] Rebase to vLLM 0.29.0 #7230 RotaryEmbedding path again on all devices. _apply_qwen_image_rotary_emb and its CPU helper unit test are removed as dead code.
• Casting cos/sin to img_query.dtype (BF16) is intentional, not a leftover bug: Qwen-Image activations are BF16, and RotaryEmbedding.forward_cuda feeds vllm_flash_attn rotary, which expects frequency tables in the same dtype as x. [Rebase] Rebase to vLLM 0.29.0 #7230 treated that round as a mismatch vs the Diffusers helper; this e2e gate measures pipeline images, where the BF16 kernel path is the one that matches.

Local L20X, same nightly recipe (FA3 hub, --fa-deterministic, 512² / 20-step / seed 42). Absolute numbers are not H100; this box is stable:

tree	SSIM	PSNR	result
#7185 038c9b948f	0.987344	34.317 dB	PASS
#7185 + only the #7230 CUDA RoPE hunk	0.948505	25.981 dB	FAIL (PSNR < 27)
revert CUDA branch to RotaryEmbedding	0.987344	34.317 dB	PASS

25.98 is the same band as H100 26.13. Not lowering PSNR_THRESHOLD.

Test plan

• Local L20X: tests/e2e/accuracy/test_qwen_image.py::test_qwen_image_matches_diffusers after revert → SSIM 0.987344 / PSNR 34.317
• Nightly H100 (Buildkite #15339, nightly-test): Diffusion X2I(&A&T) · Accuracy Test
  • test_qwen_image_matches_diffusers → SSIM 0.978375 / PSNR 31.83 (then raised SSIM_THRESHOLD to 0.97; PSNR_THRESHOLD stays 27)
  • test_diffusers_backend_t2i_matches_diffusers passed
• Ready (path-triggered, needs ready): Diffusion · Qwen Image Test
pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
• Merge (path-triggered, needs merge-test): Diffusion · Qwen Image Test
pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'advanced_model and cuda' --run-level 'advanced_model'
  Ready/merge serving jobs did not run on #15339 (nightly pipeline only). They should run when those labels are applied; this change only restores eager CUDA RoPE.

Test Result

• Local: SSIM 0.987344 / PSNR 34.317 vs the [Rebase] Rebase to vLLM 0.29.0 #7230 helper regression (SSIM 0.948505 / PSNR 25.981).
• H100 #15339 accuracy: SSIM 0.978375 / PSNR 31.83.
• Ready/merge Qwen Image serving: not executed on #15339; pending ready / merge-test CI.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/offloader.md, docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @Bounty-hunter @fhfuih @wtomin

Routing: @Bounty-hunter via module of the changed files, CODEOWNERS; @fhfuih via module of the changed files, CODEOWNERS; @wtomin via module of the changed files, CODEOWNERS

@NumberWan, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit 1ec6cc77c6ad produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[NumberWan]

@yenuo26 PTAL

[fhfuih]

I do remember having tested this code when reviewing #6110. The CUDAs-specific branch is introduced when Rebasing onto 0.29.0, and this PR actually reverts it. When testing it back then, I remember CUDA does pass the accuracy test. Together with the latest test results in the PR description, this PR looks good to me.

[yenuo26]

please fix pre-commit

[NumberWan]

please fix pre-commit

@yenuo26 Thanks — pre-commit is fixed on this branch (4ba9d93).
The GitHub pre-commit check has not gone green yet because the workflow is still waiting for a runner.

[yenuo26]

@vllm-omni-review-bot

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Direct under experiment vllm-omni-strict-5050-20260829.

[vllm-omni-review-bot]

Omni ReviewBot review

PR description

This bugfix restores Qwen-Image’s non-fused CUDA RoPE path to shared RotaryEmbedding (BF16/activation-dtype cos/sin) after #7230’s FP32 complex multiply helper regressed Omni↔Diffusers pipeline PSNR below the nightly gate. The dead _apply_qwen_image_rotary_emb helper and its unit test are removed; fused Q/K path and e2e thresholds are left intact. User-visible effect is recovering Diffusers-matching image quality on the CUDA eager RoPE path without loosening PSNR_THRESHOLD.

Change flow

flowchart TD
  A["[EXISTING] QwenImageCrossAttention.forward<br/>vid/txt freqs + qk_norm"]:::existing
  B["[CHANGED] _qwen_image_qk_norm_rope<br/>eager: RotaryEmbedding on all devices"]:::changed
  C["[REMOVED] _apply_qwen_image_rotary_emb<br/>CUDA FP32 complex multiply"]:::removed
  D["[CHANGED] benchmark + fused fallback tests<br/>native/eager pin RotaryEmbedding"]:::changed
  E["[EXISTING] test_qwen_image_matches_diffusers<br/>SSIM/PSNR vs Diffusers pipeline"]:::existing
  A --> B
  B -.-> C
  B --> D
  B --> E
  classDef existing fill:#e5e7eb,stroke:#6b7280,color:#111827
  classDef changed fill:#fef3c7,stroke:#d97706,color:#451a03,stroke-width:2px
  classDef new fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:2px
  classDef removed fill:#fee2e2,stroke:#dc2626,color:#450a0a,stroke-width:2px

Loading

See inline comments below.

[yenuo26]

LGTM

[yenuo26]

LGTM