<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>KV-Cache on 111qqz's blog</title><link>https://111qqz.com/en/tags/kv-cache/</link><description>Recent content in KV-Cache on 111qqz's blog</description><generator>Hugo -- gohugo.io</generator><language>en</language><managingEditor>hust.111qqz@gmail.com (111qqz)</managingEditor><webMaster>hust.111qqz@gmail.com (111qqz)</webMaster><copyright>© 2011-2026 111qqz</copyright><lastBuildDate>Thu, 05 Nov 2026 21:55:00 +0800</lastBuildDate><atom:link href="https://111qqz.com/en/tags/kv-cache/index.xml" rel="self" type="application/rss+xml"/><item><title>从 MHA 到 GQA：为什么现代 LLM 不再为每个 Query Head 分配独立的 KV？</title><link>https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-11-05-mha-to-gqa-kv-cache/</link><pubDate>Thu, 05 Nov 2026 21:55:00 +0800</pubDate><author>hust.111qqz@gmail.com (111qqz)</author><guid>https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-11-05-mha-to-gqa-kv-cache/</guid><description>&lt;p&gt;写完 &lt;a href="https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-09-19-multi-head-attention/" &gt;Multi-Head Attention&lt;/a&gt; 和 &lt;a href="https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-10-27-causal-attention-autoregressive-training/" &gt;Causal Attention&lt;/a&gt; 之后，我一直觉得 Attention 这条线已经差不多了——公式拆过了，routing space 的 factorization 也想明白了，Causal Mask 的因果约束也理清了。&lt;/p&gt;
&lt;p&gt;直到最近看 Llama 2 的模型配置，发现一行让我停下来的参数：&lt;/p&gt;</description></item></channel></rss>