<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Paper/Parallelism - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/paper/parallelism/</link>
        <description>Paper/Parallelism - 标签 - vllbc02&#39;s blogs</description>
        <generator>Hugo -- gohugo.io</generator><language>zh-CN</language><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Thu, 30 Jul 2026 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/paper/parallelism/" rel="self" type="application/rss+xml" /><item>
    <title>Training extremely large neural networks across thousands of GPUs</title>
    <link>https://blog.vllbc.top/distributed-training/</link>
    <pubDate>Thu, 30 Jul 2026 00:00:00 &#43;0000</pubDate>
    <author>Jeremy Jordan</author>
    <guid>https://blog.vllbc.top/distributed-training/</guid>
    <description><![CDATA[<h1 id="training-extremely-large-neural-networks-across-thousands-of-gpus">Training extremely large neural networks across thousands of GPUs</h1>
<blockquote>
<p>[!info] TL;DR
<strong>核心思想</strong>：当模型规模大到「单卡装不下、单卡跑不完」时，要在<strong>两条轴</strong>上做并行——<strong>数据轴</strong>（同一个模型复制到 N 张卡、各自跑不同 batch）和<strong>模型轴</strong>（把模型本身切碎，层间是 <strong>pipeline</strong>、层内是 <strong>tensor</strong>）。三种并行按 <strong>NVLink &gt; InfiniBand &gt; Ethernet</strong> 的网络拓扑分层叠加，是 Llama 3.1 405B 跑在 16K H100 上的根本方式。
<strong>最大收获</strong>：并行不是「GPU 越多越快」——<strong>通信原语</strong>（all-reduce / all-gather / reduce-scatter）和 <strong>gradient noise scale</strong> 才是决定 scaling 能不能继续的两道坎。
<strong>为什么重要</strong>：做 LLM 训练系统的同学必须搞清这三种并行的「同步点在哪、通信量多大、适合放在节点内还是跨节点」，否则写出来的并行策略不是慢就是 bubble 满天飞。</p>]]></description>
</item>
</channel>
</rss>
