机器之心 (微信公众号)T2·78 pts
- At 10M tokens, 98% of sparse-attention prefill time goes to the token-selection indexer rather than attention compute
- LSA delivers 22.7x single-layer prefill speedup on GLM-5.3-Flash and 72.4x on Qwen3.8-27B with near-parity accuracy
- Paper and code are both public for reproduction and integration