<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>LLM推理优化 on zzszmyf</title><link>https://doraemeng.dev/tags/llm%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96/</link><description>Recent content in LLM推理优化 on zzszmyf</description><generator>Hugo</generator><language>zh-CN</language><copyright>© 2026 zzszmyf</copyright><lastBuildDate>Thu, 29 Oct 2026 00:00:00 +0800</lastBuildDate><atom:link href="https://doraemeng.dev/tags/llm%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96/index.xml" rel="self" type="application/rss+xml"/><item><title>LLM 推理优化 · 量化（Quantization）系统学习笔记</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96-quantization%E9%87%8F%E5%8C%96%E5%AD%A6%E4%B9%A0%E7%AC%94%E8%AE%B0/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96-quantization%E9%87%8F%E5%8C%96%E5%AD%A6%E4%B9%A0%E7%AC%94%E8%AE%B0/</guid><description/></item><item><title>LLM 量化从入门到部署：11 篇精读笔记总览</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</guid><description/></item><item><title>INT8 和 FP16 有什么区别？量化前的数值编码基础</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E6%95%B0%E5%80%BC%E7%BC%96%E7%A0%81%E4%B8%8E%E8%AE%A1%E7%AE%97%E6%9C%BA%E8%A1%A8%E7%A4%BA%E5%9F%BA%E7%A1%80/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E6%95%B0%E5%80%BC%E7%BC%96%E7%A0%81%E4%B8%8E%E8%AE%A1%E7%AE%97%E6%9C%BA%E8%A1%A8%E7%A4%BA%E5%9F%BA%E7%A1%80/</guid><description/></item><item><title>量化误差是怎么算出来的？均匀量化公式推导</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-%E9%87%8F%E5%8C%96%E9%97%AE%E9%A2%98%E5%BD%A2%E5%BC%8F%E5%8C%96%E4%B8%8E%E5%9D%87%E5%8C%80%E9%87%8F%E5%8C%96%E7%90%86%E8%AE%BA/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-%E9%87%8F%E5%8C%96%E9%97%AE%E9%A2%98%E5%BD%A2%E5%BC%8F%E5%8C%96%E4%B8%8E%E5%9D%87%E5%8C%80%E9%87%8F%E5%8C%96%E7%90%86%E8%AE%BA/</guid><description/></item><item><title>FP16、BF16、FP8、FP4 有什么区别？该选哪种数值格式</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-%E6%95%B0%E5%80%BC%E6%A0%BC%E5%BC%8F%E4%B8%8E%E7%A1%AC%E4%BB%B6/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-%E6%95%B0%E5%80%BC%E6%A0%BC%E5%BC%8F%E4%B8%8E%E7%A1%AC%E4%BB%B6/</guid><description/></item><item><title>量化粒度怎么选？为什么离群值会让 INT8 精度崩掉</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-%E9%87%8F%E5%8C%96%E7%B2%92%E5%BA%A6%E6%A0%A1%E5%87%86%E4%B8%8E%E7%A6%BB%E7%BE%A4%E5%80%BC/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-%E9%87%8F%E5%8C%96%E7%B2%92%E5%BA%A6%E6%A0%A1%E5%87%86%E4%B8%8E%E7%A6%BB%E7%BE%A4%E5%80%BC/</guid><description/></item><item><title>RTN 和 GPTQ 有什么区别？权重量化方法怎么选</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96i-rtn%E4%B8%8Egptq/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96i-rtn%E4%B8%8Egptq/</guid><description/></item><item><title>AWQ 量化原理是什么？和 GPTQ、SqueezeLLM 怎么选</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96ii-awq-squeezellm-quip/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96ii-awq-squeezellm-quip/</guid><description/></item><item><title>激活量化为什么比权重量化难？LLM.int8() 与 SmoothQuant 原理</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-07-%E6%BF%80%E6%B4%BB%E9%87%8F%E5%8C%96-llm-int8%E4%B8%8Esmoothquant/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-07-%E6%BF%80%E6%B4%BB%E9%87%8F%E5%8C%96-llm-int8%E4%B8%8Esmoothquant/</guid><description/></item><item><title>KV Cache 量化怎么做？KIVI 与「量化还是压缩」的区别</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-08-kv-cache%E9%87%8F%E5%8C%96%E4%B8%8Ekivi/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-08-kv-cache%E9%87%8F%E5%8C%96%E4%B8%8Ekivi/</guid><description/></item><item><title>QAT 什么时候才值得做？PTQ、QLoRA、BitNet 路线对比</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-09-qat%E4%B8%8E%E8%AE%AD%E7%BB%83%E5%86%85%E9%87%8F%E5%8C%96-ste-qlora-bitnet/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-09-qat%E4%B8%8E%E8%AE%AD%E7%BB%83%E5%86%85%E9%87%8F%E5%8C%96-ste-qlora-bitnet/</guid><description/></item><item><title>量化后精度掉了怎么定位？perplexity 与评测怎么用</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-10-%E8%B4%A8%E9%87%8F%E8%AF%84%E4%BC%B0%E6%96%B9%E6%B3%95%E8%AE%BA/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-10-%E8%B4%A8%E9%87%8F%E8%AF%84%E4%BC%B0%E6%96%B9%E6%B3%95%E8%AE%BA/</guid><description/></item><item><title>量化模型怎么部署到 vLLM？QServe 与 FP8 Attention 实践</title><link>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-11-%E7%B3%BB%E7%BB%9F%E5%8D%8F%E5%90%8C%E4%B8%8E%E9%83%A8%E7%BD%B2/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E9%87%8F%E5%8C%96%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-11-%E7%B3%BB%E7%BB%9F%E5%8D%8F%E5%90%8C%E4%B8%8E%E9%83%A8%E7%BD%B2/</guid><description/></item><item><title>投机采样与推测解码：6 篇精读笔记总览</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</guid><description/></item><item><title>投机采样为什么是无损的？接受率与期望 token 公式推导</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E9%97%AE%E9%A2%98%E5%BD%A2%E5%BC%8F%E5%8C%96%E4%B8%8E%E6%8E%A5%E5%8F%97%E7%8E%87%E6%95%B0%E5%AD%A6/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E9%97%AE%E9%A2%98%E5%BD%A2%E5%BC%8F%E5%8C%96%E4%B8%8E%E6%8E%A5%E5%8F%97%E7%8E%87%E6%95%B0%E5%AD%A6/</guid><description/></item><item><title>投机采样原理是什么？草稿模型与拒绝采样怎么配合</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-%E5%8E%9F%E5%A7%8B%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81-%E8%8D%89%E7%A8%BF%E6%A8%A1%E5%9E%8B%E4%B8%8E%E6%8B%92%E7%BB%9D%E9%87%87%E6%A0%B7/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-%E5%8E%9F%E5%A7%8B%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81-%E8%8D%89%E7%A8%BF%E6%A8%A1%E5%9E%8B%E4%B8%8E%E6%8B%92%E7%BB%9D%E9%87%87%E6%A0%B7/</guid><description/></item><item><title>Medusa 是什么？多头解码为什么能加速 LLM 推理</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-medusa-%E5%A4%9A%E5%A4%B4%E8%A7%A3%E7%A0%81/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-medusa-%E5%A4%9A%E5%A4%B4%E8%A7%A3%E7%A0%81/</guid><description/></item><item><title>EAGLE 推测解码为什么比 Medusa 快？（含 EAGLE-2 动态草稿树）</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-eagle-%E7%89%B9%E5%BE%81%E7%A9%BA%E9%97%B4%E8%8D%89%E7%A8%BF/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-eagle-%E7%89%B9%E5%BE%81%E7%A9%BA%E9%97%B4%E8%8D%89%E7%A8%BF/</guid><description/></item><item><title>n-gram 投机采样怎么做？不用草稿模型的加速方案</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-n-gram%E6%A3%80%E7%B4%A2%E5%BC%8F%E4%B8%8E%E6%97%A0%E6%A8%A1%E5%9E%8B%E8%B7%AF%E7%BA%BF/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-n-gram%E6%A3%80%E7%B4%A2%E5%BC%8F%E4%B8%8E%E6%97%A0%E6%A8%A1%E5%9E%8B%E8%B7%AF%E7%BA%BF/</guid><description/></item><item><title>推测解码上线怎么验收？接受率与加速比怎么测</title><link>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E7%B3%BB%E7%BB%9F%E9%9B%86%E6%88%90%E4%B8%8E%E7%94%9F%E4%BA%A7%E9%AA%8C%E6%94%B6/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%8E%A8%E6%B5%8B%E8%A7%A3%E7%A0%81%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E7%B3%BB%E7%BB%9F%E9%9B%86%E6%88%90%E4%B8%8E%E7%94%9F%E4%BA%A7%E9%AA%8C%E6%94%B6/</guid><description/></item><item><title>LLM 注意力与计算内核：8 篇精读笔记总览</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-00-%E6%80%BB%E8%A7%88%E4%B8%8E%E5%AD%A6%E4%B9%A0%E5%9C%B0%E5%9B%BE/</guid><description/></item><item><title>注意力计算复杂度是怎么来的？prefill 和 decode 差在哪</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%E5%9F%BA%E7%A1%80%E4%B8%8E%E5%A4%8D%E6%9D%82%E5%BA%A6%E5%88%86%E6%9E%90/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-01-%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%E5%9F%BA%E7%A1%80%E4%B8%8E%E5%A4%8D%E6%9D%82%E5%BA%A6%E5%88%86%E6%9E%90/</guid><description/></item><item><title>FlashAttention 原理是什么？为什么又省显存又快</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-flashattention-io%E6%84%9F%E7%9F%A5%E7%9A%84%E7%B2%BE%E7%A1%AE%E6%B3%A8%E6%84%8F%E5%8A%9B/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-02-flashattention-io%E6%84%9F%E7%9F%A5%E7%9A%84%E7%B2%BE%E7%A1%AE%E6%B3%A8%E6%84%8F%E5%8A%9B/</guid><description/></item><item><title>MQA、GQA、MLA 有什么区别？KV Cache 该怎么省</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%A4%B4%E5%8F%98%E4%BD%93-mqa-gqa-mla/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-03-%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%A4%B4%E5%8F%98%E4%BD%93-mqa-gqa-mla/</guid><description/></item><item><title>稀疏注意力、滑动窗口和线性注意力怎么选</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-%E7%A8%80%E7%96%8F%E6%BB%91%E5%8A%A8%E7%AA%97%E5%8F%A3%E4%B8%8E%E7%BA%BF%E6%80%A7%E6%B3%A8%E6%84%8F%E5%8A%9B/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-04-%E7%A8%80%E7%96%8F%E6%BB%91%E5%8A%A8%E7%AA%97%E5%8F%A3%E4%B8%8E%E7%BA%BF%E6%80%A7%E6%B3%A8%E6%84%8F%E5%8A%9B/</guid><description/></item><item><title>PagedAttention 是怎么省下 KV Cache 显存的？</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-pagedattention%E4%B8%8Ekv%E6%98%BE%E5%AD%98%E7%AE%A1%E7%90%86/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-05-pagedattention%E4%B8%8Ekv%E6%98%BE%E5%AD%98%E7%AE%A1%E7%90%86/</guid><description/></item><item><title>注意力内核怎么优化？算子融合、CUDA Graph 与 Triton</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E5%86%85%E6%A0%B8%E4%BC%98%E5%8C%96%E4%B8%8E%E7%AE%97%E5%AD%90%E8%9E%8D%E5%90%88/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-06-%E5%86%85%E6%A0%B8%E4%BC%98%E5%8C%96%E4%B8%8E%E7%AE%97%E5%AD%90%E8%9E%8D%E5%90%88/</guid><description/></item><item><title>注意力优化上线怎么验收？吞吐、显存与精度怎么测</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-07-%E7%B3%BB%E7%BB%9F%E9%9B%86%E6%88%90%E4%B8%8E%E7%94%9F%E4%BA%A7%E9%AA%8C%E6%94%B6/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-07-%E7%B3%BB%E7%BB%9F%E9%9B%86%E6%88%90%E4%B8%8E%E7%94%9F%E4%BA%A7%E9%AA%8C%E6%94%B6/</guid><description/></item><item><title>SGLang RadixAttention 原理是什么？前缀缓存怎么提升命中率</title><link>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-08-%E5%89%8D%E7%BC%80%E7%BC%93%E5%AD%98%E4%B8%8Ekv%E5%A4%8D%E7%94%A8/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/llm%E6%B3%A8%E6%84%8F%E5%8A%9B%E5%86%85%E6%A0%B8%E7%B2%BE%E8%AF%BB%E7%AC%94%E8%AE%B0-08-%E5%89%8D%E7%BC%80%E7%BC%93%E5%AD%98%E4%B8%8Ekv%E5%A4%8D%E7%94%A8/</guid><description/></item><item><title>AWQ 和 GPTQ 有什么区别？4-bit 权重量化怎么选</title><link>https://doraemeng.dev/notes/awq%E5%92%8Cgptq%E6%9C%89%E4%BB%80%E4%B9%88%E5%8C%BA%E5%88%AB-4bit%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96%E6%80%8E%E4%B9%88%E9%80%89/</link><pubDate>Fri, 25 Sep 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/awq%E5%92%8Cgptq%E6%9C%89%E4%BB%80%E4%B9%88%E5%8C%BA%E5%88%AB-4bit%E6%9D%83%E9%87%8D%E9%87%8F%E5%8C%96%E6%80%8E%E4%B9%88%E9%80%89/</guid><description/></item><item><title>企业 AI 落地该选 FP8 还是 FP4？量化格式的成本、质量与硬件决策</title><link>https://doraemeng.dev/notes/%E4%BC%81%E4%B8%9Aai%E8%90%BD%E5%9C%B0%E9%80%89fp8%E8%BF%98%E6%98%AFfp4-%E9%87%8F%E5%8C%96%E6%A0%BC%E5%BC%8F%E7%9A%84%E6%88%90%E6%9C%AC%E8%B4%A8%E9%87%8F%E4%B8%8E%E7%A1%AC%E4%BB%B6%E5%86%B3%E7%AD%96/</link><pubDate>Sat, 17 Oct 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/%E4%BC%81%E4%B8%9Aai%E8%90%BD%E5%9C%B0%E9%80%89fp8%E8%BF%98%E6%98%AFfp4-%E9%87%8F%E5%8C%96%E6%A0%BC%E5%BC%8F%E7%9A%84%E6%88%90%E6%9C%AC%E8%B4%A8%E9%87%8F%E4%B8%8E%E7%A1%AC%E4%BB%B6%E5%86%B3%E7%AD%96/</guid><description/></item><item><title>vLLM、SGLang、TensorRT-LLM 该怎么选？</title><link>https://doraemeng.dev/notes/vllm-sglang-tensorrt-llm%E6%80%8E%E4%B9%88%E9%80%89/</link><pubDate>Thu, 29 Oct 2026 00:00:00 +0800</pubDate><guid>https://doraemeng.dev/notes/vllm-sglang-tensorrt-llm%E6%80%8E%E4%B9%88%E9%80%89/</guid><description/></item></channel></rss>