<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>GPU on C.CUI's Log</title><link>https://cuicaihao.github.io/zh/tags/gpu/</link><description>Recent content in GPU on C.CUI's Log</description><generator>Hugo</generator><language>zh</language><lastBuildDate>Sat, 08 Aug 2026 07:00:00 +1000</lastBuildDate><atom:link href="https://cuicaihao.github.io/zh/tags/gpu/index.xml" rel="self" type="application/rss+xml"/><item><title>优化大模型推理：PagedAttention 如何实现 GPU 显存高效管理</title><link>https://cuicaihao.github.io/zh/posts/2026-08-08-optimizing-llm-inference-the-role-of-pagedattention-in-efficient-gpu-memory-management/</link><pubDate>Sat, 08 Aug 2026 07:00:00 +1000</pubDate><guid>https://cuicaihao.github.io/zh/posts/2026-08-08-optimizing-llm-inference-the-role-of-pagedattention-in-efficient-gpu-memory-management/</guid><description>本文探讨了大型语言模型（LLM）推理中 GPU 算力利用率低下的核心问题，指出尽管 GPU 昂贵且模型庞大，但显存管理不善常导致资源闲置。文章深入介绍了 vLLM 的 PagedAttention 技术如何借鉴虚拟内存思想，结合连续批处理等方法，显著提升了 GPU 显存利用率和 LLM 推理服务的整体吞吐量。</description></item></channel></rss>