<?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>Megatron on LLM Stories</title><link>https://wgzesg.github.io/llm_stories/tags/megatron/</link><description>Recent content in Megatron on LLM Stories</description><generator>Hugo</generator><language>en</language><lastBuildDate>Wed, 29 Apr 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://wgzesg.github.io/llm_stories/tags/megatron/index.xml" rel="self" type="application/rss+xml"/><item><title>Walking Tensor Parallelism Through a Full Block</title><link>https://wgzesg.github.io/llm_stories/posts/03-tp-through-a-full-block/</link><pubDate>Wed, 29 Apr 2026 00:00:00 +0000</pubDate><guid>https://wgzesg.github.io/llm_stories/posts/03-tp-through-a-full-block/</guid><description>How to split a full transformer block across two GPUs, with concrete shapes traced through every step. Start with column-parallel everywhere, see why it costs four gathers per block, then pair it with row-parallel to land at the Megatron pattern of two all-reduces per block.</description></item></channel></rss>