PapersBuildersMulti-source10:03Chinese contentGAAC:无损 KV 压缩方案,让长上下文推理更省内存苹果和英伟达设备都能跑的无损 KV 压缩,内存省 26%、吞吐提 29%,做端侧长上下文推理的值得看看这套编解码思路。#KV compression#LLM inference#GAAC#on-device deployment9 sourcesEventaarXiv cs.AI@Yipin Guo et al. (3)SourceRead laterReadWorth followingUsefulFollow KV compression
PapersBuildersMulti-source09:59Chinese contentAcoustic-to-Text KV压缩:让全双工语音模型长对话内存降64.6%全双工语音模型聊久了内存会爆,这篇用转录旁路压缩KV,十分钟会话峰值缓存直接砍掉64.6%,性能还不掉。#MiniCPM-o 4.5#full-duplex voice#KV compression#LoRA3 sourcesEventaarXiv cs.AI@Yejin Lee et al. (4)SourceRead laterReadWorth followingUsefulFollow MiniCPM-o 4.5
Papers73°09:05Chinese contentPASK:结构化生成中的条件化KV持久化PASK让结构化生成更高效,在保持准确性的同时大幅降低内存使用,提升处理速度。#PASK#KV compression#Qwen3-4B#BFCLaarXiv cs.LG@Linze Wu, Xinrui ChenSourceRead laterReadWorth followingUsefulFollow PASK