DAMP：衰减感知的循环状态混合精度量化

AI 摘要

DAMP是一种针对Gated DeltaNet和Kimi Delta Attention等混合架构语言模型的新型量化方法，首次研究其循环状态的后训练量化。该方法通过结合量化误差能量和衰减持久性识别高风险通道，以混合精度存储，在9.9比特每状态值下保持接近FP32基线的精度，同时减少69.1%的存储并加速解码。

正文节选

DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization Abstract Softmax attention stores key and value vectors for every preceding token, causing inference memory to grow with sequence length. Recent language models incorporating Gated DeltaNet (GDN) or Kimi Delta Attention (KDA) reduce this cost by replacing the KV cache in most layers with fixed-size recurrent states. However, these recurrent states are commonly stored in FP32 and consume substantial GPU memory; their updates are memor

发布时间：2026-08-31 12:00

抓取时间：2026-08-31 12:01

来源机构：arXiv