<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
	<channel>
		<title>Fp8 on verl Blog</title>
		<link>https://verl-project.github.io/tags/fp8/</link>
		<description>Recent content in Fp8 on verl Blog</description>
		<generator>Hugo</generator>
		<language>en-us</language>
		
		
		
			<copyright>verl-project</copyright>
		
		
			<lastBuildDate>Fri, 23 Oct 2026 00:00:00 +0000</lastBuildDate>
		
			<atom:link href="https://verl-project.github.io/tags/fp8/index.xml" rel="self" type="application/rss+xml" />
			<item>
				<title>End-to-End MXFP8 RL on Blackwell: What Matching Buys, When It Pays Off, and How It Fails</title>
				<link>https://verl-project.github.io/posts/2026-10-23-mxfp8-end-to-end-rl/</link>
				<pubDate>Fri, 23 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://verl-project.github.io/posts/2026-10-23-mxfp8-end-to-end-rl/</guid>
				<description>MXFP8 on both sides of RL in verl on B200: a sampler on the learner&amp;rsquo;s weight grid cuts train-rollout KL by 16%, training pays off above about 4.5k tokens per micro-batch, and seven failures show what a sync must preserve.</description>
			</item>
	</channel>
</rss>
