开源雷达国内 GitHub 镜像站
加速通道已就绪 返回列表
全部项目

baidu-baige/LU-KV

[ICML 2026] LU-KV: a framework that optimizes head-wise KV cache budget allocation by maximizing the long-horizon marginal utility of tokens, as proposed in our paper: "Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction".

语言:Jupyter Notebook · GitHub 星标:9 · Fork:2

在 GitHub 上查看原仓库

正在读取项目信息…