---
id: 20261107-T0-05
title: "LowRAM：低秩键与摊销值实现测试时KV缓存压缩"
title_en: "LowRAM Compresses KV Cache at Test Time with Low-Rank Keys"
url: https://ai.daily.yangsir.net/daily/20261107-T0-05
issue_date: 2026-11-07
publish_date: 2026-11-06T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.07025
---

# LowRAM：低秩键与摊销值实现测试时KV缓存压缩

arXiv新论文提出LowRAM，用于测试时的KV缓存压缩。自回归解码依赖KV缓存，其内存占用随上下文长度线性增长，限制批大小和吞吐。现有后训练方法利用token、层和特征维度的冗余，或降低数值精度。LowRAM用低秩键和摊销值实现可组合的缓存压缩。对长上下文推理场景，这可在有限显存下支持更大批次，提升服务吞吐。

## English Version

**LowRAM Compresses KV Cache at Test Time with Low-Rank Keys**

A new arXiv paper introduces LowRAM for test-time KV cache compression. Autoregressive decoding relies on the KV cache, whose memory grows linearly with context length, capping batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or reduce numerical precision. LowRAM uses low-rank keys and amortized values for composable cache compression. For long-context inference, this can support larger batches under fixed GPU memory and raise serving throughput.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.07025)

**详情页**：https://ai.daily.yangsir.net/daily/20261107-T0-05

---

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