<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Prefix-Caching on Chi's Blog</title><link>https://zhangchitc.github.io/tags/prefix-caching/</link><description>Recent content in Prefix-Caching on Chi's Blog</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Thu, 05 Nov 2026 13:00:00 +0800</lastBuildDate><atom:link href="https://zhangchitc.github.io/tags/prefix-caching/index.xml" rel="self" type="application/rss+xml"/><item><title>RadixAttention: How SGLang Reuses the KV Cache Across Requests</title><link>https://zhangchitc.github.io/posts/radix-attention/</link><pubDate>Thu, 05 Nov 2026 13:00:00 +0800</pubDate><guid>https://zhangchitc.github.io/posts/radix-attention/</guid><description>SGLang keeps the KV cache in a radix tree: edges hold runs of tokens, nodes hold the GPU slot indices where their KV lives. We walk through the slot pool, TreeNode, inserting and splitting nodes, LRU eviction, and four requests that fill the pool and force an eviction.</description></item></channel></rss>