<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Lab notebook · Karthik Babu</title>
    <link>https://karthikbabu.in/notebook</link>
    <description>Field notes from hands-on R&amp;D on real hardware.</description>
    <language>en-gb</language>
    <atom:link href="https://karthikbabu.in/rss.xml" rel="self" type="application/rss+xml"/>
    
    <item>
      <title><![CDATA[Why repeated identical prompts look faster: prefix caching]]></title>
      <link>https://karthikbabu.in/notebook/2025-11-10-prefix-caching</link>
      <guid>https://karthikbabu.in/notebook/2025-11-10-prefix-caching</guid>
      <pubDate>Mon, 10 Nov 2025 00:00:00 GMT</pubDate>
      <description><![CDATA[Re-sending the same prompt can look much faster because engines reuse computed KV for a shared prefix. That is a real effect — and a trap if you ignore it in a benchmark.]]></description>
    </item>
    <item>
      <title><![CDATA[Why Docker can't use the Mac GPU]]></title>
      <link>https://karthikbabu.in/notebook/2025-11-09-docker-mac-gpu</link>
      <guid>https://karthikbabu.in/notebook/2025-11-09-docker-mac-gpu</guid>
      <pubDate>Sun, 09 Nov 2025 00:00:00 GMT</pubDate>
      <description><![CDATA[On Apple Silicon, Docker runs Linux guests. Those guests do not get Metal, so GPU-backed inference has to run on the host.]]></description>
    </item>
    <item>
      <title><![CDATA[Running vLLM on a Mac]]></title>
      <link>https://karthikbabu.in/notebook/2025-11-08-vllm-on-mac</link>
      <guid>https://karthikbabu.in/notebook/2025-11-08-vllm-on-mac</guid>
      <pubDate>Sat, 08 Nov 2025 00:00:00 GMT</pubDate>
      <description><![CDATA[I got vLLM running on Apple Silicon with the Metal plugin and measured it against plain Transformers on the same model and questions.]]></description>
    </item>
  </channel>
</rss>