
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>HANS</title>
      <link>https://www.hansking.cn/blog</link>
      <description>A posts created with Next.js and Tailwind.css</description>
      <language>en-us</language>
      <managingEditor>admin@hansking.cn (HANS)</managingEditor>
      <webMaster>admin@hansking.cn (HANS)</webMaster>
      <lastBuildDate>Sun, 14 Sep 2025 10:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.hansking.cn/tags/PagedAttention/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.hansking.cn/blog/2026/vLLM-tui-li-fu-wu-PagedAttention-yu-tun-tu-you-hua</guid>
    <title>vLLM 推理服务：PagedAttention 与吞吐优化</title>
    <link>https://www.hansking.cn/blog/2026/vLLM-tui-li-fu-wu-PagedAttention-yu-tun-tu-you-hua</link>
    <description>[{&quot;value&quot;:&quot;vLLM 推理服务：PagedAttention 与吞吐优化&quot;,&quot;url&quot;:&quot;#vllm-推理服务pagedattention-与吞吐优化&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先用最小服务确认链路&quot;,&quot;url&quot;:&quot;#一先用最小服务确认链路&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、PagedAttention 解决什么问题&quot;,&quot;url&quot;:&quot;#二pagedattention-解决什么问题&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、动态批处理让请求不必互相等待&quot;,&quot;url&quot;:&quot;#三动态批处理让请求不必互相等待&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、max-model-len 不是越大越好&quot;,&quot;url&quot;:&quot;#四max-model-len-不是越大越好&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、流式输出要正确处理断开和取消&quot;,&quot;url&quot;:&quot;#五流式输出要正确处理断开和取消&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、请求调度要考虑公平性&quot;,&quot;url&quot;:&quot;#六请求调度要考虑公平性&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、监控不能只有 GPU 利用率&quot;,&quot;url&quot;:&quot;#七监控不能只有-gpu-利用率&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、压测要使用真实长度分布&quot;,&quot;url&quot;:&quot;#八压测要使用真实长度分布&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、常见优化顺序&quot;,&quot;url&quot;:&quot;#九常见优化顺序&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十、发布和回滚要一起设计&quot;,&quot;url&quot;:&quot;#十发布和回滚要一起设计&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：推理服务的核心是调度资源&quot;,&quot;url&quot;:&quot;#总结推理服务的核心是调度资源&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 14 Sep 2025 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>vLLM</category><category>PagedAttention</category><category>推理服务</category>
  </item>

    </channel>
  </rss>
