<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>prλtam — Pritam Ravi</title>
    <link>https://pritamravi.com</link>
    <description>Pritam Ravi explores inference engineering, ML systems, GPU performance and creative intelligence. Building systems and learning from first principles.</description>
    <language>en</language>
    <lastBuildDate>Sun, 01 Nov 2026 09:00:00 +0000</lastBuildDate>
    <atom:link href="https://pritamravi.com/feed.xml" rel="self" type="application/rss+xml"/>
    <item>
      <title>Anatomy of a decode step</title>
      <link>https://pritamravi.com/writing/anatomy-of-a-decode-step/</link>
      <guid isPermaLink="true">https://pritamravi.com/writing/anatomy-of-a-decode-step/</guid>
      <pubDate>Wed, 21 Oct 2026 09:00:00 +0000</pubDate>
      <description><![CDATA[One token out, an entire system in. Tracing a single autoregressive decode step from the scheduler&#39;s queue to a sampled logit — and finding out where the time actually goes.]]></description>
      <category>gpu</category>    </item>
    <item>
      <title>The KV cache is an allocator wearing an ML costume</title>
      <link>https://pritamravi.com/writing/the-kv-cache-is-an-allocator/</link>
      <guid isPermaLink="true">https://pritamravi.com/writing/the-kv-cache-is-an-allocator/</guid>
      <pubDate>Mon, 14 Sep 2026 09:00:00 +0000</pubDate>
      <description><![CDATA[Autoregressive inference spends most of its intelligence not on matrix multiplication but on remembering. A first-principles look at why the KV cache behaves like a memory allocator — and why that framing explains almost everything about serving performance.]]></description>
      <category>systems</category>    </item>
  </channel>
</rss>