arXiv · cs.AI · Kunming Shao, Xijie Huang, Jierun Chen, Jiaqi Shao, Yanli Wang, Jiangnan Yu, Qiwei Li, Wei Zhang, Kwang-Ting Cheng, Chi Ying Tsui · 5 小时前模型评分41
FlagRot：面向低比特 KV 缓存的选择性旋转方法
FlagRot: Selective Rotation for Low-Bit KV Caches
导读
FlagRot 提出按 32 通道分片决定是否旋转 key，读取时改为旋转 query，从而避免旋转对平坦 key 引入新峰值。在 2.5 比特下，它把 KIVI 的十二模型 KL 散度从 18.7 降至 12.4×10^-3，并在 9/12 模型上优于 RotateKV。
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
- “We therefore propose FlagRot, which stores each 32-channel piece rotated or as is, whichever has the smaller calibrated attention error, and rotates the query instead when reading.”
- “Served as a vLLM plugin, it raises RULER at 128k tokens by +2.9 points and the LongBench v1 mean by +0.4 to +1.9 points on the three of four models where 2-bit KIVI loses accuracy.”
来源：arXiv · cs.AI · arxiv.org