<?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>Quantization on Bits-Entangled</title><link>https://stondo.github.io/tags/quantization/</link><description>Recent content in Quantization on Bits-Entangled</description><generator>Hugo -- 0.157.0</generator><language>en-us</language><lastBuildDate>Tue, 25 Aug 2026 12:00:00 +0000</lastBuildDate><atom:link href="https://stondo.github.io/tags/quantization/index.xml" rel="self" type="application/rss+xml"/><item><title>Three Days, Four Wrong Hypotheses, and One Uncensored 27B That Finally Serves on a Single RTX 5090</title><link>https://stondo.github.io/posts/obliteratus-qwen38-27b-nvfp4-rtx5090/</link><pubDate>Tue, 25 Aug 2026 12:00:00 +0000</pubDate><guid>https://stondo.github.io/posts/obliteratus-qwen38-27b-nvfp4-rtx5090/</guid><description>The full story of quantizing OBLITERATUS/Qwen3.8-27B-OBLITERATED to NVFP4 for a single RTX 5090 with vLLM and LMCache: every ModelOpt trap I fell into, the assertion that survived three fixes, and the one debugging trick that cracked it in a single run.</description></item></channel></rss>