[Revert] drop Wan2.2 prompt-length enforcement from #2847

[david6666666]

Summary

• raise the default Wan2.2 max_sequence_length from 512 to 2048
• apply the same default across Wan2.2 T2V, I2V, TI2V, and VACE pipelines
• avoid padding short prompts and negative prompts to the configured ceiling during text encoding
• add regression coverage for both prompt-length validation and actual-length text-encoder padding

Root cause

Issue #2874 is triggered because Wan2.2 pipelines defaulted max_sequence_length to 512, while the CI I2V prompt is 654 tokens long.

After raising the limit, encode_prompt() still used padding="max_length" with max_length=max_sequence_length, which padded short prompts to 2048 and introduced an unnecessary text-encoding slowdown for short-prompt requests.

Validation

• python -m py_compile 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 tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py
• python -m pytest --noconftest -q tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py
• pre-commit run --all-files
• Wan2.2 I2V E2E with the local snapshot at /mnt/data1/huggingface/hub/models--Wan-AI--Wan2.2-I2V-A14B-Diffusers/snapshots/596658fd9ca6b7b71d5057529bbf319ecbc61d74
  • short prompt request completed successfully
  • inference_time_s=114.31679305434227
  • profiler showed Wan22I2VPipeline.text_encoder.forward back at roughly 0.015s per short-prompt call

[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 ISSUE: This PR changes the default max_sequence_length from 512 to 2048, which is a breaking change that affects all Wan2.2 users.

Issue #2874 is a CI test failure. The appropriate fix is to either:

1. Shorten the CI test prompt to fit within 512 tokens, OR
2. Add explicit configuration (max_sequence_length=2048) to the test setup

Changing the default value increases memory usage for ALL users just to fix a CI test. This is not the right approach.

Please reconsider the fix for #2874 without breaking existing user behavior.

[hsliuustc0106]

Raising max_sequence_length to 2048 is a 4x increase. This fixes the immediate CI failure (#2874), but consider:

• Does this match Wan2.2's actual trained prompt length limit? If the model was trained on 512-token prompts, going to 2048 could produce degraded quality or unexpected behavior beyond the training distribution.

• Memory impact: longer prompts = more VRAM for text encoder outputs. For I2V workflows where VRAM is already tight, this could cause OOMs on smaller GPUs.

• For future PRs: consider documenting the model's native token limit in model metadata rather than hardcoding in the pipeline. This makes it easier to adjust per-deployment without code changes.

[Gaohan123]

Plan to temporarily skip it with #2883 after this one is updated

[david6666666]

Follow-up update for the short-prompt regression introduced after raising Wan2.2 max_sequence_length.

• Added commit b665b917 ([Fix] avoid padding short Wan2.2 prompts to max_sequence_length)
• Updated Wan2.2 T2V, I2V, and TI2V so they still validate against the configured max_sequence_length, but only pad text-encoder inputs to the actual max prompt length needed by the current batch
• Added regression coverage in tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py

Validation run on this branch:

• python -m py_compile ... passed
• python -m pytest --noconftest -q tests/diffusion/models/wan2_2/test_wan22_max_sequence_length.py passed (16 passed)
• pre-commit run --all-files passed
• Wan2.2 I2V E2E with the short English prompt completed successfully
  • inference_time_s=114.31679305434227
  • profiler showed Wan22I2VPipeline.text_encoder.forward back at roughly 0.015s per short-prompt call

This keeps the larger prompt-length support for #2874 while removing the short-prompt text-encoding regression.

[david6666666]

@gcanlin @fan2956 wan2.2 i2v accuracy test is failed

[david6666666]

Branch update:

• Reset this PR branch onto the latest upstream/main
• Reverted only the Wan2.2-related changes that came from #2847
• Kept the shared vllm_omni/diffusion/utils/prompt_utils.py because it is now used by Qwen-Image paths as well

Current branch head:

• 9528def2 [Revert] drop Wan2.2 prompt-length enforcement from #2847

Validation on this branch:

• python -m py_compile for the reverted Wan2.2 pipeline files passed
• python -m pytest --noconftest -q tests/diffusion/models/wan2_2/test_wan22_i2v_pipeline.py tests/diffusion/models/wan2_2/test_wan22_pipeline_diffuse.py tests/diffusion/models/wan2_2/test_wan22_pipeline_helpers.py tests/diffusion/models/wan2_2/test_wan22_ti2v_pipeline.py tests/diffusion/models/wan2_2/test_wan22_vace_pipeline.py
  • 16 passed
• pre-commit run --all-files
  • Passed

So the branch now contains only the targeted Wan2.2 revert on top of the current main.

[gcanlin]

LGTM. Please also revert 5be6ff5#diff-f53575bf88041d823a9739163c467042dace2b532fc93a825f9f3e89fc169315 this change in v0.18.0.post1.