<?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>Inference on 安橙的博客</title><link>https://blog.ans20xx.com/tags/inference/</link><description>Recent content in Inference on 安橙的博客</description><generator>Hugo -- 0.163.3</generator><language>zh</language><lastBuildDate>Sat, 20 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://blog.ans20xx.com/tags/inference/index.xml" rel="self" type="application/rss+xml"/><item><title>Day 29 · LLM 推理基础</title><link>https://blog.ans20xx.com/posts/ai/day29/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day29/</guid><description>进入 LLM 推理 Infra:理解 prefill 与 decode 的阶段差异、KV Cache 为什么是显存大头、吞吐/延迟指标如何拆解,并写出一个最小 generation loop。</description></item><item><title>Day 30 · 解码算法</title><link>https://blog.ans20xx.com/posts/ai/day30/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day30/</guid><description>进入 LLM 解码策略:理解 greedy、beam search、temperature、top-k、top-p 的采样语义与服务化影响;掌握 speculative decoding 的 draft/verify 思想,并用 Transformers 与 vLLM 参数做小型实验。</description></item><item><title>Day 31 · PagedAttention &amp; vLLM</title><link>https://blog.ans20xx.com/posts/ai/day31/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day31/</guid><description>学习 PagedAttention 与 vLLM 的核心机制:为什么 KV Cache 会浪费显存,如何用 block table 管理逻辑块到物理块的映射,copy-on-write 如何支撑并行采样和 beam search,以及这些机制如何服务高吞吐 LLM serving。</description></item><item><title>Day 33 · Continuous Batching</title><link>https://blog.ans20xx.com/posts/ai/day33/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day33/</guid><description>学习 LLM 推理服务中的 Continuous Batching:理解静态 batching 与 in-flight batching 的差异,prefill/decode 如何混排,以及 TGI、vLLM、SGLang 调度器在吞吐、TTFT、TPOT 与公平性上的取舍。</description></item><item><title>Day 32 · vLLM 实战</title><link>https://blog.ans20xx.com/posts/ai/day32/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0800</pubDate><guid>https://blog.ans20xx.com/posts/ai/day32/</guid><description>动手部署一个 7B 模型到 vLLM,开启 OpenAI 兼容 API,学习 --max-num-seqs 与 --gpu-memory-utilization 的调参方法,并建立推理服务压测与排错流程。</description></item><item><title>Day 34 · SGLang &amp; RadixAttention</title><link>https://blog.ans20xx.com/posts/ai/day34/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0800</pubDate><guid>https://blog.ans20xx.com/posts/ai/day34/</guid><description>学习 SGLang 推理框架与 RadixAttention:理解前缀缓存、共享 prefix 的请求调度、Radix Tree 如何复用 KV Cache,并动手用同一个 system prompt 发多请求观察缓存命中。</description></item><item><title>Day 35 · 量化 (1)：权重量化</title><link>https://blog.ans20xx.com/posts/ai/day35/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0800</pubDate><guid>https://blog.ans20xx.com/posts/ai/day35/</guid><description>学习 LLM 推理中的权重量化:理解 INT8 / INT4、per-channel / group-wise scale、GPTQ 与 AWQ 的核心思想,并用 AutoGPTQ 或 llama.cpp 完成一次模型量化与评估。</description></item></channel></rss>