---
id: 20261017-T0-09
title: "RoLaQ：为MLA的2比特KV缓存做旋转与范围联合校准"
title_en: "RoLaQ: Joint Rotation and Range Calibration for 2-Bit KV Caches in MLA"
url: https://ai.daily.yangsir.net/daily/20261017-T0-09
issue_date: 2026-10-17
publish_date: 2026-10-16T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.20686
---

# RoLaQ：为MLA的2比特KV缓存做旋转与范围联合校准

多头潜在注意力（MLA）将KV缓存内容压缩为每个token的共享潜在表示，使得低比特存储对长上下文推理很有吸引力。但在2比特精度下，表示误差会同时扰动注意力匹配和值信息。RoLaQ提出联合旋转与范围校准方法，针对MLA架构下的2比特KV缓存进行优化，目标是在极低比特下保持推理质量。该研究以预印本形式发布在arXiv上，针对的是长上下文推理中显存占用的核心瓶颈。

## English Version

**RoLaQ: Joint Rotation and Range Calibration for 2-Bit KV Caches in MLA**

Multi-head latent attention (MLA) compresses KV cache content into a shared latent per token, making low-bit storage attractive for long-context inference. But at 2 bits, representation errors perturb both attention matching and value information. RoLaQ proposes joint rotation and range calibration for 2-bit KV caches in MLA, aiming to preserve inference quality at extremely low bit widths. The work is available as an arXiv preprint, targeting the core memory bottleneck in long-context inference.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.20686)

**详情页**：https://ai.daily.yangsir.net/daily/20261017-T0-09

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*