[Tracking Issue][Help Wanted]: FlashInfer backend improvements

[comaniac]

This issue tracks the progress and roadmap of integrating new FlashInfer kernels and enabling more vLLM features in FlashInfer backend. Items listed in Milestone 1 is a part of re-arch #8779 .

If you want any features available in FlashInfer backend but not listed here, or if you're interested in taking items, please feel free to comment.

Milestone 1

• [ ] Unify FlashInfer backend with multi-level cascade attention kernel. #8132
• [ ] Support in-batch prefix sharing with 2-level cascade attention. @raywanb
• [ ] Support prefix caching.

Milestone 2

• [ ] Support chunked prefill.
• [ ] Support FP8 E4M3 kv-cache with kv scale.
• [ ] Support batch expansion in speculative decoding.
• [ ] Integrate RaggedTensor kernel.

cc @yzh119 @LiuXiaoxuanPKU @raywanb @simon-mo @WoosukKwon @pavanimajety

[pavanimajety]

Milestone 2 [x] Support chunked prefill. #9781 [x] Support FP8 E4M3 kv-cache with kv scale. #9861 [ ] Integrate RaggedTensor kernel. - In progress

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions[bot]]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!