研究库 论文知识库

内容库 / 论文

arXiv:2609.13285 · LLM 基础设施

Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction

Grouped Value Attention：通过按需 Key 重建实现高效 KV 缓存

Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction

类型：arxiv
标识：2609.13285
链接：https://arxiv.org/abs/2609.13285
主分类：llm-infra
形态：position
TLDR：The KV cache is a primary bottleneck for Transformer decoding: its memory footprint and cache-read traffic grow with sequence length. Grouped-query attention (GQA) reduces this cost by sharing key-value heads, but still stores both a key and a value at every step. We introduce Grouped Value Attention (GVA), which stores grouped values and reconstructs content keys with a learned linear map. At inference, the map can be absorbed into the query, eliminating the need to materialize content keys in the intended decode path. A small shared decoupled RoPE channel retains positional information throu
待LLM分类：否
标题中文：Grouped Value Attention：通过按需 Key 重建实现高效 KV 缓存
TLDR中文：KV 缓存是 Transformer 解码的主要瓶颈：其内存占用与缓存读取流量随序列长度增长。Grouped-query attention (GQA) 通过共享 key-value 头来降低成本，但仍在每一步同时存储 key 和 value。我们提出 Grouped Value Attention (GVA)，存储分组 value 并通过学习的线性映射重建内容 key。推理时，该映射可吸收进 query，无需在目标解码路径中实例化内容 key。一个小型共享解耦 RoPE 通道保留位置信息 throu
来源文件：
/inbox/tom/_candidates/2026-09-15-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记 PagedAttention 2.0 推理优化 · 2026-07-28

攻略循环不是「藏 token」——Looped Transformer 原理与 Astra 架构传言纠偏 · 干货攻略

笔记 Tom 文献雷达 · AI Agent / RAG / 长上下文 · 2026-07-11 下午版

仓库 fla-org/flash-linear-attention

笔记 ReMemR1: Look Back to Reason Forward (Revisitable Memory for Long-Context LLM Agents)

攻略 FlashAttention-4 · 干货攻略

arXiv

2609.13285

TLDR

KV 缓存是 Transformer 解码的主要瓶颈：其内存占用与缓存读取流量随序列长度增长。Grouped-query attention (GQA) 通过共享 key-value 头来降低成本，但仍在每一步同时存储 key 和 value。我们提出 Grouped Value Attention (GVA)，存储分组 value 并通过学习的线性映射重建内容 key。推理时，该映射可吸收进 query，无需在目标解码路径中实例化内容 key。一个小型共享解耦 RoPE 通道保留位置信息 throuThe KV cache is a primary bottleneck for Transformer decoding: its memory footprint and cache-read traffic grow with sequence length. Grouped-query attention (GQA) reduces this cost by sharing key-value heads, but still stores both a key and a value at every step. We introduce Grouped Value Attention (GVA), which stores grouped values and reconstructs content keys with a learned linear map. At inference, the map can be absorbed into the query, eliminating the need to materialize content keys in the intended decode path. A small shared decoupled RoPE channel retains positional information throu

arXiv ↗ 中文全文