<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Albert&apos;s Tech Blog</title><link>https://qiguanping.github.io/</link><description>AI Infra · High-Performance Networking · RDMA</description><item><title>DualPath 深度解读：当“空闲的另一半网卡”成为破局点</title><link>https://qiguanping.github.io/posts/dualpath/</link><description>从双路径 KV-Cache 加载、CNIC 流量管理到自适应调度，理解推理系统如何把闲置网络带宽重新变成有效吞吐。</description><pubDate>Sun, 26 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/dualpath/</guid></item><item><title>Mooncake 深度解读：用存储换算力——KVCache 中心化解耦架构的工程哲学</title><link>https://qiguanping.github.io/posts/mooncake/</link><description>把 KV Cache 从推理副产品提升为系统调度中枢，连接分布式存储、传输引擎与 P/D 分离。</description><pubDate>Sat, 25 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/mooncake/</guid></item><item><title>ZCube 深度解读：当自动化搜索推翻拓扑设计者的直觉</title><link>https://qiguanping.github.io/posts/zcube/</link><description>从 ATOP 搜索空间到低直径递归拓扑，分析性能、网络成本与适用边界之间的真实取舍。</description><pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/zcube/</guid></item><item><title>NCCLX：把通信从 GPU 里“搬”出来</title><link>https://qiguanping.github.io/posts/ncclx/</link><description>Meta 如何用 Host-driven、Zero-copy 与端网协同，在大规模 RoCE 网络上重新设计集合通信。</description><pubDate>Thu, 23 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/ncclx/</guid></item><item><title>DeepSeek-V3 硬件架构反思深度解读：在受管制硬件上榨取一流效率</title><link>https://qiguanping.github.io/posts/deepseek-v3-hardware/</link><description>从 H800、NVLink、MoE 到多平面 Fat-Tree，理解模型、通信库和网络拓扑的协同设计。</description><pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/deepseek-v3-hardware/</guid></item><item><title>Aegis 深度解读：当“不能碰客户代码”成为诊断系统的第一性原理</title><link>https://qiguanping.github.io/posts/aegis/</link><description>从日志与网管系统到定制 CCL，分析大规模训练服务如何定位故障、性能退化与交付前风险。</description><pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate><guid>https://qiguanping.github.io/posts/aegis/</guid></item></channel></rss>