[BugFix][Attention] Fix sliding window attention in V1 giving incorrect results

[LucasWilkinson]

The FA3 update to use a AOT scheduler (#13111) did not properly handle sliding window attention. FIX #17476

Tested using:

python -m pytest -vs tests/v1/e2e/test_correctness_sliding_window.py

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[Copilot]

Pull Request Overview

This PR fixes an issue where the FA3 update did not properly handle sliding window attention by ensuring a consistent sliding window configuration is used with the AOT scheduler. Key changes include:

• Introducing a new helper function (_get_sliding_window_configs) to extract sliding window configurations from the model.
• Updating the FlashAttentionMetadataBuilder to initialize and propagate a uniform sliding window configuration.
• Modifying the scheduling lambda to pass the sliding window configuration to the underlying flash attention kernel.

Comments suppressed due to low confidence (1)

vllm/v1/attention/backends/flash_attn.py:284

• FlashAttentionImpl is referenced here but is not imported in this file. Consider adding an import for FlashAttentionImpl to ensure the assertion works correctly.

assert isinstance(layer.impl, FlashAttentionImpl)

[heheda12345]

getattr(self.hf_text_config, "sliding_window", None) checks whether this model only contains sliding window layers
getattr(hf_text_config, "interleaved_sliding_window", None) is not None checks whether this model contains both sliding window and non-sliding window.
Hope these two APIs can help to make this PR more clean.

[LucasWilkinson]

getattr(self.hf_text_config, "sliding_window", None) checks whether this model only contains sliding window layers getattr(hf_text_config, "interleaved_sliding_window", None) is not None checks whether this model contains both sliding window and non-sliding window. Hope these two APIs can help to make this PR more clean.

Thanks for the pointer! Ya I thought about that, but:

    def get_hf_config_sliding_window(
            self) -> Union[Optional[int], list[Optional[int]]]

can return a list of list[Optional[int]], and was sure how to handle that. So I decided to with this route so that way its a guaranteed source of truth, and we are robust against new models using something other than sliding_window or interleaved_sliding_window (or not respecting the precedent you mentioned).

[heheda12345]

I've confirmed that it solves my problem in #17476.

I can approve it as a quick bug fix, but want to do some refactor after #17483 which moves the initialization of attention metadata builder after model loading.

[LucasWilkinson]

I can approve it as a quick bug fix, but want to do some refactor after #17483 which moves the initialization of attention metadata builder after model loading.

agreed, I think we should also unify with infer_global_hyperparameters in FlashInfer