作者位次 #2/6ACL 2026Findings· Findings of ACL
Accelerating Prefilling via Decoding-time Contribution Sparsity
Zhiyuan He · Yike Zhang · Chengruidong Zhang · Huiqiang Jiang · Yuqing Yang · Lili Qiu
摘要
Large Language Models (LLMs) incur quadratic attention complexity with input length, creating a major time bottleneck in the prefilling stage. Existing acceleration methods largely exploit attention score sparsity by estimating blocks with high attention scores and applying dynamic sparse attenti...