<?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>SGLang on Jiaqi</title>
    <link>https://blog.jiaqiguo.xyz/tags/sglang/</link>
    <description>Recent content in SGLang on Jiaqi</description>
    <generator>Hugo</generator>
    <language>zh-cn</language>
    <lastBuildDate>Fri, 24 Jul 2026 15:30:00 +0800</lastBuildDate>
    <atom:link href="https://blog.jiaqiguo.xyz/tags/sglang/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>从 4 bit 权重到高并发 API：深入浅出 GPTQ、AWQ、safetensors、vLLM 与 SGLang</title>
      <link>https://blog.jiaqiguo.xyz/posts/gpu-llm-serving-stack/</link>
      <pubDate>Fri, 24 Jul 2026 15:30:00 +0800</pubDate>
      <guid>https://blog.jiaqiguo.xyz/posts/gpu-llm-serving-stack/</guid>
      <description>下载一个带 AWQ 或 GPTQ 后缀的模型，里面为什么还是 safetensors？vLLM 与 SGLang 又在优化什么？从 W4A16 的显存账本出发，拆开 GPTQ 的逐列误差补偿、AWQ 的激活感知缩放、safetensors 的文件结构，再沿着加载、内核、KV Cache 和调度一路走到高并发推理服务。</description>
    </item>
  </channel>
</rss>
