<?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>Rl on 柏炎的技术笔记</title><link>https://iseekyan.github.io/zh/tags/rl/</link><description>Recent content in Rl on 柏炎的技术笔记</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Mon, 18 May 2026 21:50:00 +0800</lastBuildDate><atom:link href="https://iseekyan.github.io/zh/tags/rl/index.xml" rel="self" type="application/rss+xml"/><item><title>让长序列 MoE RL 训练更好调：Megatron-Lite / bumblebee 的优化实践</title><link>https://iseekyan.github.io/zh/posts/qwen35-long-sequence-moe-rl/</link><pubDate>Mon, 18 May 2026 21:50:00 +0800</pubDate><guid>https://iseekyan.github.io/zh/posts/qwen35-long-sequence-moe-rl/</guid><description>&lt;p>长序列训练不只出现在 pretrain。进入 RL 阶段以后，prompt、rollout、reward、工具调用和多轮交互都会把上下文长度拉高。很多团队关心的问题也随之变化：他们需要用较少启动卡数，稳定跑通 128K、256K，甚至更长上下文的算法实验。&lt;/p></description></item></channel></rss>