arXiv · cs.AI · Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel · 3 小时前模型评分41
DeepSeek Sparse Attention 新方法 Multi-layer Indexer：跨层复用选择提升解码速度
Multi-layer Indexer: Overlapping Selection and Attention in DeepSeek Sparse Attention
导读
针对 DeepSeek Sparse Attention（DSA，DeepSeek 稀疏注意力）中闪电索引器扫描拖慢长上下文解码的问题，研究者提出多层索引器（MLI，Multi-layer Indexer），让每层直接复用上一层索引器的选择结果，无需重训练和新权重。
分类置信度
中 · 待人工复核
- 只有单一媒体来源，需要人工复核。
来源级别：媒体与个人
事件类型：研究发布
证据摘录
- “The resulting Multi-layer Indexer (MLI) needs no retraining and no new weights.”
- “Its default realization changes no kernel, running inside SGLang's production decode graph; an optional fused decode kernel accelerates it further on DeepSeek-V3.2.”
- “We serve DeepSeek-V3.2 and GLM-5 under IndexCache's protocol at retentions 1/2 to 1/32 (the fraction of layers that keep an indexer) and contexts 10K to 195K.”
- “Its LongBench, RULER and LongBench v2 accuracy stays within a few percentage points of IndexCache's.”
来源：arXiv · cs.AI · arxiv.org