[Bugfix] enforce max_sequence_length for Qwen-Image and Wan2.2 series before encoding

[david6666666]

Summary

• enforce max_sequence_length before the text encoder runs instead of relying on silent truncation or post-encoder slicing
• make the default max_sequence_length=1024 effective across Qwen-Image, Qwen-Image-Layered, Qwen-Image-Edit, and Qwen-Image-Edit-Plus
• validate Qwen edit/edit-plus on text prompt length before image token expansion so short edit prompts still work
• make the Wan2.2 family (T2V / I2V / TI2V / VACE) reject overlong prompts before UMT5 encoding, while preserving the existing default limit of 512
• add targeted tests for both Qwen-Image and Wan2.2 prompt-length enforcement and default propagation

Validation

• python -m pytest -q tests/diffusion/models/qwen_image/test_qwen_image_max_sequence_length.py tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py
• python -m ruff check vllm_omni/diffusion/models/qwen_image/prompt_utils.py vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_edit.py vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_edit_plus.py vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_layered.py tests/diffusion/models/qwen_image/test_qwen_image_max_sequence_length.py vllm_omni/diffusion/models/wan2_2/prompt_utils.py vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_i2v.py vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_ti2v.py vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_vace.py tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py
• E2E serve + curl with local snapshots:
  • Qwen-Image: short prompt succeeded at /v1/images/generations; a 5000-word prompt failed with got 5006 tokens, but \max_sequence_length` is 1024`
  • Qwen-Image-Edit: short prompt succeeded at /v1/images/edits; a 5000-word prompt failed with got 5009 tokens, but \max_sequence_length` is 1024`
  • Wan2.2-T2V-A14B-Diffusers: a num_inference_steps=1, num_frames=5 request completed successfully at /v1/videos; a 5000-word prompt failed with got 5001 tokens, but \max_sequence_length` is 512`

Closes #2794.

[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. [Gate Failure] pre-commit - The pre-commit check is failing. Please fix any formatting/linting issues.

---

VERDICT: REQUEST_CHANGES (gate must pass before code review)

Please fix the failing pre-commit check before proceeding with the review.

[david6666666]

Update: this PR now also covers the Wan2.2 family.

Added in the second patch:

• Wan2.2-T2V / I2V / TI2V / VACE now validate the real prompt token length before UMT5 encoding instead of silently truncating
• request paths now consistently fall back to the Wan default max_sequence_length=512
• targeted Wan2.2 tests were added alongside the earlier Qwen-Image tests

Additional validation:

• python -m pytest -q tests/diffusion/models/qwen_image/test_qwen_image_max_sequence_length.py tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py
• live serve + curl on local Wan2.2-T2V-A14B-Diffusers snapshot:
  • short prompt completed successfully at /v1/videos with num_inference_steps=1, num_frames=5
  • 5000-word prompt failed with: got 5001 tokens, but max_sequence_length is 512

[lishunyang12]

Looks good overall. The approach of validating before encoding (rather than relying on silent truncation) is the right call. The shared validate_prompt_sequence_lengths utility is clean and well-documented. Tests cover all pipeline variants and both default/explicit limit paths.

A few minor observations:

1. length_offset parameter is unused. validate_prompt_sequence_lengths accepts length_offset: int = 0 but no caller ever passes it. If there is no planned use, consider removing it to keep the API surface minimal. Not blocking.

2. Double tokenization in Qwen pipelines. The PR tokenizes the prompt once with truncation=False for validation, then the text encoder runs the same text through the tokenizer again internally (or the processor does for edit pipelines). This is a minor perf cost (extra tokenizer forward pass) but acceptable for correctness. If this ever becomes a bottleneck, the validated tokens could be reused directly.

3. Qwen _get_qwen_prompt_embeds still slices [:max_sequence_length] after encoding (in encode_prompt for pipeline_qwen_image.py and pipeline_qwen_image_layered.py). Since validation now rejects overlong prompts, this slice is effectively a no-op for user text. However, the template suffix tokens can push the total sequence beyond max_sequence_length, so the post-encoding slice still serves a purpose for trimming template overhead. This is correct but worth a brief inline comment explaining why the slice is still needed.

4. Test coverage is solid. The _RejectingTextEncoder pattern is a nice way to assert the encoder is never reached for rejected prompts. The boundary tests for template suffix exclusion and image placeholder exclusion are well thought out.

LGTM. Approving.

[gcanlin]

This UT looks missing cpu and core_model pymark.

[gcanlin]

Ditto.

[gcanlin]

LGTM