<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Speculative-Decoding on 111qqz的小窝</title><link>https://111qqz.com/tags/speculative-decoding/</link><description>Recent content in Speculative-Decoding on 111qqz的小窝</description><generator>Hugo -- gohugo.io</generator><language>zh-cn</language><managingEditor>hust.111qqz@gmail.com (111qqz)</managingEditor><webMaster>hust.111qqz@gmail.com (111qqz)</webMaster><copyright>© 2011-2026 111qqz</copyright><lastBuildDate>Tue, 10 Nov 2026 20:00:00 +0800</lastBuildDate><atom:link href="https://111qqz.com/tags/speculative-decoding/index.xml" rel="self" type="application/rss+xml"/><item><title>从投机解码到 Rand5 模拟 Rand7：跨越十年的拒绝采样直觉</title><link>https://111qqz.com/2026/11/speculative-decoding-rejection-sampling/</link><pubDate>Tue, 10 Nov 2026 20:00:00 +0800</pubDate><author>hust.111qqz@gmail.com (111qqz)</author><guid>https://111qqz.com/2026/11/speculative-decoding-rejection-sampling/</guid><description>&lt;p&gt;&lt;a href="https://111qqz.com/2026/11/llm-inference-stateful-generation/" &gt;上一篇&lt;/a&gt;理清了 Prefill 和 Decode 的因果链之后，我开始翻 Decode 阶段的加速方案。结论已经很清楚了：在单请求、小 Batch 的典型场景下，Decode 阶段是 memory-bound 的——每生成一个 token 就要把整个模型权重从显存搬一遍，计算单元大部分时间在等数据（Batch 较大或模型结构不同时，瓶颈也可能转向 compute-bound，但单次请求的延迟优化场景下 memory-bound 是常态）。&lt;/p&gt;</description></item></channel></rss>