<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>SGLang on Yiwen Cai</title><link>https://yiwen-cai.github.io/tags/sglang/</link><description>Recent content in SGLang on Yiwen Cai</description><generator>Hugo -- gohugo.io</generator><language>zh-cn</language><managingEditor>caiyiwen.cs@foxmail.com (Yiwen Cai)</managingEditor><webMaster>caiyiwen.cs@foxmail.com (Yiwen Cai)</webMaster><copyright>© 2026 Yiwen Cai</copyright><lastBuildDate>Tue, 03 Nov 2026 16:17:09 +0800</lastBuildDate><atom:link href="https://yiwen-cai.github.io/tags/sglang/index.xml" rel="self" type="application/rss+xml"/><item><title>SGLang 架构精读（二）：SGLang-Omni 七阶段异构调度、流式硬实时声码器与原生全双工状态机</title><link>https://yiwen-cai.github.io/notes/llm-inference/sglang-omni-full-duplex/</link><pubDate>Tue, 03 Nov 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid>https://yiwen-cai.github.io/notes/llm-inference/sglang-omni-full-duplex/</guid><description>深入解析 SGLang-Omni 服务 Qwen3-Omni 与 MiniCPM-o 全双工多模态模型的系统架构（Ch 16–21）：涵盖 80ms 音频帧与 1+15 双层 RVQ 生成、七阶段三类异构调度器、Talker 禁用 RadixCache 与 Overlap Scheduling 的底层原因、CommRouter 四级通信传输层、多模型共卡的 85.7 GB 显存墙与 86% SM 争用博弈、Code2Wav 25+10 左上下文滑动窗口消除接缝爆音，以及 /v1/realtime 原生全双工的 SessionController 跨步状态保持。</description></item><item><title>SGLang 架构精读（三）：DP Attention、Two-Batch Overlap、HiCache 与 Mooncake PD 分离全景</title><link>https://yiwen-cai.github.io/notes/llm-inference/sglang-distributed-and-disaggregation/</link><pubDate>Tue, 03 Nov 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid>https://yiwen-cai.github.io/notes/llm-inference/sglang-distributed-and-disaggregation/</guid><description>结合 SGLang 核心源码完整补齐《Learning SGLang》Part IV–V 进阶与多卡解耦架构（Ch 22–29）：涵盖 W4A16 与 W8A8 (FP8) 量化的 Roofline 改写算术、MLA 遇上传统 TP 的 8 倍 KV 冗余灾难与 DP Attention + MoE EP 解耦破局、Two-Batch Overlap (TBO) 双微批计算/通信双流乒乓隐藏、HiCache 三级存储金字塔逐层流水预取、Mooncake GPUDirect RDMA 逐层零拷贝 PD 分离，以及 sgl-router 影子前缀树动态负载均衡。</description></item><item><title>SGLang 架构精读（一）：从三级页表、Piecewise CUDA Graph 到 MLA 矩阵吸收与投机解码</title><link>https://yiwen-cai.github.io/notes/llm-inference/sglang-single-gpu-engine/</link><pubDate>Tue, 03 Nov 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid>https://yiwen-cai.github.io/notes/llm-inference/sglang-single-gpu-engine/</guid><description>以单步耗时方程 t_step = max(t_cpu, t_comm, max(t_mem, t_compute)) 与 H100 上 Llama-3-8B 的硬件账本为主线，系统拆解 SGLang 单卡引擎与编译器栈（Ch 01–15）：涵盖 page_size=1 三级页表、RadixCache 末位盲区、Overlap Scheduling 双缓冲闭环、Piecewise CUDA Graph 拆图算术、Dynamo/Inductor 融合、DeepSeek MLA 矩阵吸收（128 FLOP/byte）及三代投机解码（EAGLE-3 / DFlash / DSpark）。</description></item></channel></rss>