[V1] Implement Cascade Attention

[WoosukKwon]

This PR implements a simple version of Cascade Attention. Cascade attention can save the HBM bandwidth for reading KV cache when requests share the same prefix.

NOTE: For simplicity, this PR only uses cascade attention when every running request shares the same KV cache. If one or more requests do not share the KV cache, cascade attention is not used.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[comaniac]

Overall LGTM. The implementation is pretty clean. One optimization we can do in the future is having a wrapper CUDA kernel in flash attention so that we can save kernel invocation overheads. This is how FlashInfer does as well.

@raywanb since you've worked on FlashInfer cascade kernel integration in v0, can you also take a look at the current v1 interface and see if this interface can integrate FlashInfer easily in the future?

[raywanb]

Overall LGTM. The implementation is pretty clean. One optimization we can do in the future is having a wrapper CUDA kernel in flash attention so that we can save kernel invocation overheads. This is how FlashInfer does as well.

@raywanb since you've worked on FlashInfer cascade kernel integration in v0, can you also take a look at the current v1 interface and see if this interface can integrate FlashInfer easily in the future?

Yes, seems pretty straightforward to integrate Flashinfer!

[comaniac]

LGTM. Just two minor questions.
Also do you have a unit test to verify the correctness?

[comaniac]

Thanks for adding this! I'll probably try to benchmark this feature next week if I got a chance.