<?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>本地部署 on kwelldo 的小家</title><link>https://blog.kwelldo.cn/tags/%E6%9C%AC%E5%9C%B0%E9%83%A8%E7%BD%B2/</link><description>Recent content in 本地部署 on kwelldo 的小家</description><generator>Hugo -- 0.168.0</generator><language>zh-cn</language><lastBuildDate>Thu, 15 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://blog.kwelldo.cn/tags/%E6%9C%AC%E5%9C%B0%E9%83%A8%E7%BD%B2/index.xml" rel="self" type="application/rss+xml"/><item><title>自我改进」改的是哪一层，决定你按什么标准选基座</title><link>https://blog.kwelldo.cn/posts/%E8%87%AA%E6%88%91%E6%94%B9%E8%BF%9B%E6%94%B9%E7%9A%84%E6%98%AF%E5%93%AA%E4%B8%80%E5%B1%82-%E6%88%91%E7%BB%99agent%E9%80%89%E5%9F%BA%E5%BA%A7%E6%9C%80%E5%90%8E%E6%B2%A1%E9%80%89%E6%9C%AC%E5%9C%B0/</link><pubDate>Thu, 15 Oct 2026 00:00:00 +0000</pubDate><guid>https://blog.kwelldo.cn/posts/%E8%87%AA%E6%88%91%E6%94%B9%E8%BF%9B%E6%94%B9%E7%9A%84%E6%98%AF%E5%93%AA%E4%B8%80%E5%B1%82-%E6%88%91%E7%BB%99agent%E9%80%89%E5%9F%BA%E5%BA%A7%E6%9C%80%E5%90%8E%E6%B2%A1%E9%80%89%E6%9C%AC%E5%9C%B0/</guid><description>给自部署 Agent 选基座，第一问不是&amp;#34;哪个模型可训练&amp;#34;，而是&amp;#34;你的 Agent 到底改哪一层&amp;#34;。改配置层，选型只看 JSON 可靠性与成本；本地部署的辩护从省钱收缩到自主，连同前提一起留档延期。</description></item><item><title>官方说的「稀疏注意力」，在我机器上根本拿不到——一次视频生成加速的三档实测</title><link>https://blog.kwelldo.cn/posts/%E5%AE%98%E6%96%B9%E8%AF%B4%E7%9A%84%E7%A8%80%E7%96%8F%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%9C%A8%E6%88%91%E6%9C%BA%E5%99%A8%E4%B8%8A%E6%A0%B9%E6%9C%AC%E6%8B%BF%E4%B8%8D%E5%88%B0-%E4%B8%80%E6%AC%A1%E8%A7%86%E9%A2%91%E7%94%9F%E6%88%90%E5%8A%A0%E9%80%9F%E7%9A%84%E4%B8%89%E6%A1%A3%E5%AE%9E%E6%B5%8B/</link><pubDate>Sat, 10 Oct 2026 00:00:00 +0000</pubDate><guid>https://blog.kwelldo.cn/posts/%E5%AE%98%E6%96%B9%E8%AF%B4%E7%9A%84%E7%A8%80%E7%96%8F%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%9C%A8%E6%88%91%E6%9C%BA%E5%99%A8%E4%B8%8A%E6%A0%B9%E6%9C%AC%E6%8B%BF%E4%B8%8D%E5%88%B0-%E4%B8%80%E6%AC%A1%E8%A7%86%E9%A2%91%E7%94%9F%E6%88%90%E5%8A%A0%E9%80%9F%E7%9A%84%E4%B8%89%E6%A1%A3%E5%AE%9E%E6%B5%8B/</guid><description>我在本地用 ComfyUI 跑视频生成，一镜到底 10 秒，20 步基线一次要 370 秒。为了提速，我去下了一个模型卡上标着「85% 稀疏注意力」的加速版。跑完三档对照，速度确实上来了：20 步 370 秒，8 步 1</description></item><item><title>35B 不等于 35B 显存——我只看文件大小下模型，装完才发现装不下</title><link>https://blog.kwelldo.cn/posts/35b%E4%B8%8D%E7%AD%89%E4%BA%8E35b%E6%98%BE%E5%AD%98-%E6%88%91%E5%8F%AA%E7%9C%8B%E6%96%87%E4%BB%B6%E5%A4%A7%E5%B0%8F%E4%B8%8B%E6%A8%A1%E5%9E%8B%E8%A3%85%E5%AE%8C%E6%89%8D%E5%8F%91%E7%8E%B0%E8%A3%85%E4%B8%8D%E4%B8%8B/</link><pubDate>Thu, 08 Oct 2026 00:00:00 +0000</pubDate><guid>https://blog.kwelldo.cn/posts/35b%E4%B8%8D%E7%AD%89%E4%BA%8E35b%E6%98%BE%E5%AD%98-%E6%88%91%E5%8F%AA%E7%9C%8B%E6%96%87%E4%BB%B6%E5%A4%A7%E5%B0%8F%E4%B8%8B%E6%A8%A1%E5%9E%8B%E8%A3%85%E5%AE%8C%E6%89%8D%E5%8F%91%E7%8E%B0%E8%A3%85%E4%B8%8D%E4%B8%8B/</guid><description>上次我在本机部署 Qwen3.8-27B，前后搭进去一晚上。下载 13 个 G，配 systemd，写 jinja 模板，接进 agent。全部弄完，重启服务，日志里跳出一行：显存不足，OOM。</description></item></channel></rss>