---
id: 20260922-T0-08
title: "长文本解码显存吃紧？学习型稀疏注意力按需加载KV缓存"
title_en: "Elastic Threshold Attention Cuts KV Cache Bottleneck in Long-Context Decoding"
url: https://ai.daily.yangsir.net/daily/20260922-T0-08
issue_date: 2026-09-22
publish_date: 2026-09-21T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.20888
---

# 长文本解码显存吃紧？学习型稀疏注意力按需加载KV缓存

针对长文本解码时KV缓存造成的显存带宽瓶颈，研究者提出Elastic Threshold Attention（弹性阈值注意力）。现有稀疏注意力方法靠固定启发式规则选择性加载KV，容易丢弃必要上下文，导致生成质量下降。该方法改为学习上下文相关的稀疏模式，按内容动态决定哪些KV需要加载。论文目前挂在arXiv上（编号2609.20888），尚未公布与密集注意力的具体质量对比数据。若效果成立，长文本推理的显存占用和延迟有望同时下降。

## English Version

**Elastic Threshold Attention Cuts KV Cache Bottleneck in Long-Context Decoding**

Researchers propose Elastic Threshold Attention to address memory-bandwidth bottlenecks caused by massive KV caches during long-context decoding. Existing sparse attention methods rely on rigid heuristics to selectively load KV pairs, which can drop necessary context and degrade output quality. This approach instead learns contextual sparsity, dynamically deciding which KV entries to load based on content. Posted on arXiv (2609.20888), the paper has not yet released quality comparisons against dense attention. If validated, it could reduce both memory footprint and latency for long-context inference.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.20888)

**详情页**：https://ai.daily.yangsir.net/daily/20260922-T0-08

---

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