<?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>Flash Attention on Vlad Savinov</title>
    <link>https://vladsavinov.com/tags/flash-attention/</link>
    <description>Recent content in Flash Attention on Vlad Savinov</description>
    <image>
      <title>Vlad Savinov</title>
      <url>https://vladsavinov.com/images/avatar.jpg</url>
      <link>https://vladsavinov.com/images/avatar.jpg</link>
    </image>
    <generator>Hugo -- 0.146.0</generator>
    <language>en-us</language>
    <lastBuildDate>Tue, 01 Sep 2026 11:09:21 +0100</lastBuildDate>
    <atom:link href="https://vladsavinov.com/tags/flash-attention/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Another View on Context Parallel: Balancing Compute</title>
      <link>https://vladsavinov.com/posts/another-view-on-context-parallel-balancing-compute/</link>
      <pubDate>Tue, 01 Sep 2026 00:00:00 +0000</pubDate>
      <guid>https://vladsavinov.com/posts/another-view-on-context-parallel-balancing-compute/</guid>
      <description>Why context parallelism can improve training throughput by reducing variance in Flash Attention execution time across data-parallel groups.</description>
    </item>
  </channel>
</rss>
