STEPQuant：Delta-Rule 循环状态的时空量化
原标题：Paper page - STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization
AI 摘要
该论文提出 STEPQuant，一个针对 Delta-rule 循环状态（线性注意力中替代 KV 缓存的固定大小状态）的时空后训练量化框架。作者发现量化误差的影响在时间维度（长寿命记忆误差跨多步解码持续）和空间维度（不同 key 行对输出影响不同）上存在差异，据此按误差幅度与记忆寿命分配精度，并联合拟合 key 行与 value 列缩放。在 Qwen3.8-27B 和 Kimi-Linear-48B-A3B-Instruct 上，6-bit 配置接近 FP32 精度，4-bit 配置优于均匀 INT8；集成到 SGLang 后实现超 5 倍循环状态压缩，总服务内存最多降低 68.7%。
正文节选
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization Abstract Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complemen