VAR-Q：面向视觉自回归生成的免调优KV缓存量化
原标题：VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation
AI 摘要
该论文提出 VAR-Q，一种无需调优的 KV 缓存量化框架，用于视觉自回归图像与视频生成。作者分析发现，视觉自回归模型的 KV 缓存中不到 1% 的离群通道承载了主要功能，且这些离群值沿通道和块两个轴呈结构化分布，源于投影偏置与 RoPE 散射。VAR-Q 沿这两个轴分组量化，无需校准或重训练，在九个模型上将 KV 内存最多减少 78.5%，质量损失可忽略，并使单卡推理批次增大至 6 倍。
正文节选
VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation Abstract Visual autoregressive generation spans two dominant paradigms: next-scale prediction for images and video, and next-frame prediction for video. Both decode block-by-block rather than token-by-token, yielding KV caches whose statistics diverge from those of language models. At scale, KV memory exceeds the weights and limits large-scale deployment caused by a large amount of visual tokens. Our an