<?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>VLLM on AI 每日一课</title><link>https://ai.zlfun.cn/tags/vllm/</link><description>Recent content in VLLM on AI 每日一课</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Tue, 03 Nov 2026 10:06:15 +0800</lastBuildDate><atom:link href="https://ai.zlfun.cn/tags/vllm/index.xml" rel="self" type="application/rss+xml"/><item><title>GPT-6 Astra 不训练直接改权重：手把手复现「权重空间编程」修复版 Qwen3.5-0.8B（附 vLLM 对比）</title><link>https://ai.zlfun.cn/posts/20261103-weight-space-programming-qwen35/</link><pubDate>Tue, 03 Nov 2026 10:06:15 +0800</pubDate><guid>https://ai.zlfun.cn/posts/20261103-weight-space-programming-qwen35/</guid><description>MIT 研究者让前沿大模型直接读写神经网络权重，只改 0.016% 参数就大幅缓解了 Qwen3.5-0.8B 的思考死循环。本文教你用官方 diff 逐位重建修复版、校验改动，并用 vLLM 对比原版与修复版。</description></item></channel></rss>