<?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>inference on Margin Notes</title><link>https://www.191196.xyz/tags/inference/</link><description>Recent content in inference on Margin Notes</description><generator>Hugo</generator><language>en-US</language><lastBuildDate>Fri, 30 Oct 2026 09:26:00 +0000</lastBuildDate><atom:link href="https://www.191196.xyz/tags/inference/index.xml" rel="self" type="application/rss+xml"/><item><title>Three papers I keep re-reading about inference cost</title><link>https://www.191196.xyz/posts/inference-papers/</link><pubDate>Fri, 30 Oct 2026 09:26:00 +0000</pubDate><guid>https://www.191196.xyz/posts/inference-papers/</guid><description>Paged attention, speculative decoding, and 4-bit weights — what each one actually buys.</description></item><item><title>What actually fits on an 8 GB card</title><link>https://www.191196.xyz/posts/local-8gb/</link><pubDate>Wed, 22 Jul 2026 20:41:00 +0000</pubDate><guid>https://www.191196.xyz/posts/local-8gb/</guid><description>The weights fit. The context is what gets you.</description></item></channel></rss>