[diffusion][Wan]: fix sparse attention backends being applied to cross-attention

[Ratish1]

Motivation

This PR ensures that sparse attention backends (such as sliding_tile_attn and video_sparse_attn) are restricted to self-attention layers in Wan models. Fixes #17303

Modifications

• Added backend filtering to WanT2VCrossAttention and WanI2VCrossAttention constructors to force a fallback to standard backends for cross-attention.
• Updated CausalWanTransformerBlock to correctly pass supported backends to its cross-attention layer.

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @Ratish1, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request rectifies a critical architectural oversight in Wan models where specialized sparse attention backends, intended solely for self-attention mechanisms, were being erroneously applied to cross-attention layers. By introducing explicit filtering at the constructor level for cross-attention modules and ensuring proper backend propagation, the change guarantees that attention mechanisms operate as designed, preventing potential miscomputations and ensuring model stability and correctness.

Highlights

• Fix for Sparse Attention Backends: Addressed an issue where sparse attention backends were incorrectly applied to cross-attention layers in Wan models, ensuring they are restricted to self-attention layers only. This resolves issue [Bug][Diffusion][Wan] STA (and other sparse attention) is applied to cross‑attention in Wan; should be self‑attn‑only? #17303.
• Cross-Attention Backend Filtering: Implemented explicit filtering of sparse attention backends within the constructors of WanT2VCrossAttention and WanI2VCrossAttention to prevent their use in cross-attention contexts.
• CausalWanTransformerBlock Update: Modified CausalWanTransformerBlock to correctly pass the supported_attention_backends argument to its cross-attention layer, enabling the new filtering logic.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request effectively addresses the issue of sparse attention backends being incorrectly applied to cross-attention layers in Wan models. The changes ensure that WanT2VCrossAttention and WanI2VCrossAttention correctly filter out sparse attention backends, falling back to standard backends as intended. The update to CausalWanTransformerBlock to pass the supported_attention_backends parameter is also appropriate. Overall, the changes align with the motivation to restrict sparse attention to self-attention layers.

[adarshxs]

/tag-and-rerun-ci

[adarshxs]

cc @mickqian this LGTM

[Ratish1]

Hey @mickqian , does this PR require more changes?. Thanks

[Ratish1]

I updated the PR @mickqian , lmk if you need more changes. Thanks

[Ratish1]

CI passes @mickqian . Thanks.

[tie-pilot-qxw]

It seems like this pull request missed handling the sparse video generation 2 backend. Specifically, the logic for removing the svg2 backend from cross_attn_backends appears to have been overridden. This can lead to incorrect execution due to improper handling of the backends. Could you please address this? Thank you!

[Ratish1]

Hey @tie-pilot-qxw , thanks for pointing this out. I opened a PR here #18900. Thanks.