RecSA
by Hao AI Lab · Technique · Text & code
Sparse attention searching a tree over the KV cache; selection cost and entries read per query grow as O(log L)
Evidence
- Nov 17Code released with kernels, Qwen3 integration, RULER evaluation and latency benchmarks
- Nov 17Initialized from Qwen3-4B and trained on 0.62B tokens at 32K, scores 60.5 on RULER at 1M versus 40.0 for the best baseline (Dense + DCA)
- Nov 17On one H200, prefill is 3.2x and decoding 3.8x faster than dense FlashAttention at 524K; decoding stays near 6 ms from 131K to 524K