<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <entry>
    <title>MIG 与动态 TP：GPU 切卡与张量并行弹性化的两条路</title>
    <link href="https://1193700079.github.io/mig-dynamic-tp.html"/>
    <id>https://1193700079.github.io/mig-dynamic-tp.html</id>
    <published>2026-11-16T11:00:00.000Z</published>
    <updated>2026-11-16T11:00:00.000Z</updated>
    <summary type="html">&lt;p&gt;阅读 MIG 硬件切卡与动态 TP 两条路线，解释为什么 RTX PRO 6000 这类工作站卡都不能运行时重组张量并行。&lt;/p&gt;</summary>
    <category term="技术阅读" scheme="https://1193700079.github.io/categories/%E6%8A%80%E6%9C%AF%E9%98%85%E8%AF%BB/"/>
    <category term="llm" scheme="https://1193700079.github.io/tags/llm/"/>
  </entry>
  <entry>
    <title>Reinforcement Learning-driven Diffusion Models：语言、视觉与更远的边界</title>
    <link href="https://1193700079.github.io/rl-diffusion-survey.html"/>
    <id>https://1193700079.github.io/rl-diffusion-survey.html</id>
    <published>2026-11-16T10:00:00.000Z</published>
    <updated>2026-11-16T10:00:00.000Z</updated>
    <summary type="html">&lt;p&gt;RL 驱动的扩散模型系统综述：技术全景、四大领域、三大瓶颈。TPAMI 接收。&lt;/p&gt;</summary>
    <category term="技术阅读" scheme="https://1193700079.github.io/categories/%E6%8A%80%E6%9C%AF%E9%98%85%E8%AF%BB/"/>
    <category term="llm" scheme="https://1193700079.github.io/tags/llm/"/>
  </entry>
  <entry>
    <title>LeapQuant：近乎无损的 8-bit 线性注意力递归态量化</title>
    <link href="https://1193700079.github.io/leapquant.html"/>
    <id>https://1193700079.github.io/leapquant.html</id>
    <published>2026-11-16T09:00:00.000Z</published>
    <updated>2026-11-16T09:00:00.000Z</updated>
    <summary type="html">&lt;p&gt;三个免训练技巧把线性注意力递归态量化到近乎无损的 8-bit。&lt;/p&gt;</summary>
    <category term="技术阅读" scheme="https://1193700079.github.io/categories/%E6%8A%80%E6%9C%AF%E9%98%85%E8%AF%BB/"/>
    <category term="llm" scheme="https://1193700079.github.io/tags/llm/"/>
  </entry>
  <entry>
    <title>Databricks：如何让 Qwen 3.8 27B 微调快6倍</title>
    <link href="https://1193700079.github.io/qwen-sft-6x.html"/>
    <id>https://1193700079.github.io/qwen-sft-6x.html</id>
    <published>2026-11-16T08:00:00.000Z</published>
    <updated>2026-11-16T08:00:00.000Z</updated>
    <summary type="html">&lt;p&gt;GPU 平均利用率已经 98%，训练却还能快 6 倍。Databricks 完整记录了排查过程、算子变化与数据打包细节。&lt;/p&gt;</summary>
    <category term="技术阅读" scheme="https://1193700079.github.io/categories/%E6%8A%80%E6%9C%AF%E9%98%85%E8%AF%BB/"/>
    <category term="llm" scheme="https://1193700079.github.io/tags/llm/"/>
  </entry>
  <title>卿人 Blog</title>
  
  
  <link href="https://1193700079.github.io/atom.xml" rel="self"/>
  
  <link href="https://1193700079.github.io/"/>
  <updated>2021-12-28T14:09:38.858Z</updated>
  <id>https://1193700079.github.io/</id>
  
  <author>
    <name>RuiQing Yang</name>
    
  </author>
  
  <generator uri="https://hexo.io/">Hexo</generator>
  
  
  
  
  
  
  
  
  
  
  
</feed>
