Congratulations on the MLSys acceptance! I really enjoyed reading your paper and found the Match–Amend–Complete idea very interesting.
I have a question about Table 3: End-to-end Attention Latency. In the table, Quest and RocketKV appear to have higher latency than the FlashInfer full-attention baseline, even under very small KV budgets. This is an important and interesting result, and I would like to better understand the measurement setup.
Would it be possible to share the code or scripts used to reproduce the latency results in Table 3, especially the Quest-related evaluation? I am particularly interested in:
- How the end-to-end attention latency was measured.
- Whether the reported latency includes the full Quest pipeline, such as page/token selection, metadata processing, KV gathering, and the final attention computation?
- Whether the Quest implementation was based on the official repository and whether any modifications were made to integrate it with SGLang / FlashInfer?
Understanding this would be very helpful for interpreting the system-level comparison in Table 3.
Thanks a lot, and congratulations again on the great work!
Congratulations on the MLSys acceptance! I really enjoyed reading your paper and found the Match–Amend–Complete idea very interesting.
I have a question about Table 3: End-to-end Attention Latency. In the table, Quest and RocketKV appear to have higher latency than the FlashInfer full-attention baseline, even under very small KV budgets. This is an important and interesting result, and I would like to better understand the measurement setup.
Would it be possible to share the code or scripts used to reproduce the latency results in Table 3, especially the Quest-related evaluation? I am particularly interested in:
Understanding this would be very helpful for interpreting the system-level comparison in Table 3.
Thanks a lot, and congratulations again on the great work!