<?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>PagedAttention on Ringi's Log</title><link>https://lilinji.github.io/tags/pagedattention/</link><description>Recent content in PagedAttention on Ringi's Log</description><generator>Hugo -- 0.152.2</generator><language>zh-cn</language><lastBuildDate>Tue, 01 Sep 2026 00:00:00 +0800</lastBuildDate><atom:link href="https://lilinji.github.io/tags/pagedattention/index.xml" rel="self" type="application/rss+xml"/><item><title>第08讲：为什么 Attention 算力没少算，FlashAttention 却能快 3 倍？——从 Decoder Block 算子图、RoPE 复数旋转到显存四账本的第一性原理</title><link>https://lilinji.github.io/2026/09/%E7%AC%AC08%E8%AE%B2%E4%B8%BA%E4%BB%80%E4%B9%88-attention-%E7%AE%97%E5%8A%9B%E6%B2%A1%E5%B0%91%E7%AE%97flashattention-%E5%8D%B4%E8%83%BD%E5%BF%AB-3-%E5%80%8D%E4%BB%8E-decoder-block-%E7%AE%97%E5%AD%90%E5%9B%BErope-%E5%A4%8D%E6%95%B0%E6%97%8B%E8%BD%AC%E5%88%B0%E6%98%BE%E5%AD%98%E5%9B%9B%E8%B4%A6%E6%9C%AC%E7%9A%84%E7%AC%AC%E4%B8%80%E6%80%A7%E5%8E%9F%E7%90%86/</link><pubDate>Tue, 01 Sep 2026 00:00:00 +0800</pubDate><guid>https://lilinji.github.io/2026/09/%E7%AC%AC08%E8%AE%B2%E4%B8%BA%E4%BB%80%E4%B9%88-attention-%E7%AE%97%E5%8A%9B%E6%B2%A1%E5%B0%91%E7%AE%97flashattention-%E5%8D%B4%E8%83%BD%E5%BF%AB-3-%E5%80%8D%E4%BB%8E-decoder-block-%E7%AE%97%E5%AD%90%E5%9B%BErope-%E5%A4%8D%E6%95%B0%E6%97%8B%E8%BD%AC%E5%88%B0%E6%98%BE%E5%AD%98%E5%9B%9B%E8%B4%A6%E6%9C%AC%E7%9A%84%E7%AC%AC%E4%B8%80%E6%80%A7%E5%8E%9F%E7%90%86/</guid><description>深入剖析 Transformer 算子执行图与第一性原理：以认知冲突拆解 Self-Attention、FlashAttention 片上 SRAM 算子融合、RoPE 2D 复数几何旋转、SwiGLU 门控、RMSNorm 梯度高速公路与 LLaMA 7B/70B 显存四账本及 PagedAttention 虚拟内存分页实战。</description></item></channel></rss>