本文在 KV cache 约束下对 LLM 推理进行理论建模,提出一种新型批处理与调度算法,在有效管理 KV cache 内存的同时最小化推理延迟,并通过在合成数据集上与后视最优的对比展示其强劲的实证性能。This work model LLM inference with KV cache constraints theoretically and proposes a novel batching and scheduling algorithm that minimizes inference latency while effectively managing the KV cache's memory, and demonstrates the algorithm's strong empirical performance by comparing it to the hindsight optimal in a synthetic dataset.
论文
1 张论文卡片 · LLM 基础设施 · 评测集
14. LLM 推理在线调度:hindsight optimal benchmark