跳到正文
arXiv Artificial Intelligence· OneRec Team, Chenglong Chu, Guohong Mu, Guowang Zhang, Hao Peng, Hui Wang, Jian Liang, Jiangxia Cao, Jiao Ou, Jiaxin Deng, Lu Ren, Muhao Wei, Ruiming Tang, Tianle Xu, Wenwu Ou, Xingmei Wang, Zhixin Ling, Zhiyuan Liang, Ziqi Wang, Chaoyi Ma, Haixiao Gao, Kuo Cai, Renpu Liu, Ruochen Yang, Wenlong Xu, Xinqi Jin, Yan Sun, Yifan Wang, Yingzhi He, Yiyu Wang, Yusen Huo, Zhaojie Liu, Zihan Zhou·· 18 小时前AI 评分27

快手 KSA V2 技术报告:让 Summary token 兼顾 KV 内存与稀疏索引

Kwai Summary Attention V2 Technical Report

AI 导读

快手 OneRec 团队提出 Kwai Summary Attention V2(KSA V2),针对混合 KSA-Full 架构中摘要压缩块跨层必须一致、Full attention 层仍付二次预填充开销两个瓶颈,让 Summary token 同时承担 KV 内存与稀疏索引两种角色,从而实现更灵活的 KV Cache 压缩并降低 Full attention 层计算成本。

来源:arXiv Artificial Intelligence · arxiv.org