📦 GitHub: https://github.com/geekchow/deepseek-v41-flash-inference-guide
你在哪里 :这是整套指南的入口。
读完本文你将知道:这次发布包含哪些东西(一份技术报告 + 两个开源组件),整套指南怎么读,以及所有核心概念和组件的全景图。
1. 一句话说清这次发布
2026 年 10 月 20 日,DeepSeek 在知乎发布技术报告《DeepSeek V4.1 Flash 大规模推理部署实践》,讲清楚了支撑 V4.1 Flash 线上服务的整套推理系统;同时开源了两个生产级组件:
| 发布物 | 是什么 | 仓库 / 链接 | 许可证 |
|---|---|---|---|
| 技术报告 | V4.1 Flash 推理系统的设计目标、整体架构与关键实现(含未开源的 KV 缓存和负载均衡设计) | 知乎原文;配套 Inference Deployment Q&A | --- |
| DeepInfer | 大规模云服务推理框架(推理引擎 + 高性能算子),支撑 V4.1 Flash 线上服务 | github.com/deepseek-ai/DeepInfer | Apache-2.0 |
| deepseek-recipe | Token 编解码组件:把主流 LLM API 协议与模型 token 序列互相转换,行为与官方 API 完全对齐 | github.com/deepseek-ai/deepseek-recipe | MIT |
标题数字 (官方 Q&A):单集群每天服务 155 万亿 token ,整体收入 38 美元 / GPU·小时(按谷时价格计)。
2. 全景思维导图(所有概念与组件)
这张图回答:这套系统解决什么问题、由哪些概念和组件构成、我们用哪个例子把它们串起来。
#mermaid-svg-O0tQQpKs83eiyBID{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-O0tQQpKs83eiyBID .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-O0tQQpKs83eiyBID .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-O0tQQpKs83eiyBID .error-icon{fill:#552222;}#mermaid-svg-O0tQQpKs83eiyBID .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-O0tQQpKs83eiyBID .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-O0tQQpKs83eiyBID .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-O0tQQpKs83eiyBID .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-O0tQQpKs83eiyBID .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-O0tQQpKs83eiyBID .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-O0tQQpKs83eiyBID .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-O0tQQpKs83eiyBID .marker{fill:#333333;stroke:#333333;}#mermaid-svg-O0tQQpKs83eiyBID .marker.cross{stroke:#333333;}#mermaid-svg-O0tQQpKs83eiyBID svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-O0tQQpKs83eiyBID p{margin:0;}#mermaid-svg-O0tQQpKs83eiyBID .edge{stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .section--1 rect,#mermaid-svg-O0tQQpKs83eiyBID .section--1 path,#mermaid-svg-O0tQQpKs83eiyBID .section--1 circle,#mermaid-svg-O0tQQpKs83eiyBID .section--1 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section--1 path{fill:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section--1 text{fill:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon--1{font-size:40px;color:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge--1{stroke:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth--1{stroke-width:17;}#mermaid-svg-O0tQQpKs83eiyBID .section--1 line{stroke:hsl(60, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-0 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-0 path,#mermaid-svg-O0tQQpKs83eiyBID .section-0 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-0 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-0 path{fill:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-O0tQQpKs83eiyBID .section-0 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-0{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-0{stroke:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-0{stroke-width:14;}#mermaid-svg-O0tQQpKs83eiyBID .section-0 line{stroke:hsl(240, 100%, 83.5294117647%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-1 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-1 path,#mermaid-svg-O0tQQpKs83eiyBID .section-1 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-1 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-1 path{fill:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-1 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-1{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-1{stroke:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-1{stroke-width:11;}#mermaid-svg-O0tQQpKs83eiyBID .section-1 line{stroke:hsl(260, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-2 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-2 path,#mermaid-svg-O0tQQpKs83eiyBID .section-2 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-2 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-2 path{fill:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-2 text{fill:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-2{font-size:40px;color:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-2{stroke:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-2{stroke-width:8;}#mermaid-svg-O0tQQpKs83eiyBID .section-2 line{stroke:hsl(90, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-3 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-3 path,#mermaid-svg-O0tQQpKs83eiyBID .section-3 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-3 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-3 path{fill:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-3 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-3{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-3{stroke:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-3{stroke-width:5;}#mermaid-svg-O0tQQpKs83eiyBID .section-3 line{stroke:hsl(120, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-4 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-4 path,#mermaid-svg-O0tQQpKs83eiyBID .section-4 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-4 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-4 path{fill:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-4 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-4{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-4{stroke:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-4{stroke-width:2;}#mermaid-svg-O0tQQpKs83eiyBID .section-4 line{stroke:hsl(150, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-5 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-5 path,#mermaid-svg-O0tQQpKs83eiyBID .section-5 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-5 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-5 path{fill:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-5 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-5{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-5{stroke:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-5{stroke-width:-1;}#mermaid-svg-O0tQQpKs83eiyBID .section-5 line{stroke:hsl(180, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-6 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-6 path,#mermaid-svg-O0tQQpKs83eiyBID .section-6 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-6 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-6 path{fill:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-6 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-6{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-6{stroke:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-6{stroke-width:-4;}#mermaid-svg-O0tQQpKs83eiyBID .section-6 line{stroke:hsl(210, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-7 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-7 path,#mermaid-svg-O0tQQpKs83eiyBID .section-7 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-7 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-7 path{fill:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-7 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-7{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-7{stroke:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-7{stroke-width:-7;}#mermaid-svg-O0tQQpKs83eiyBID .section-7 line{stroke:hsl(270, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-8 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-8 path,#mermaid-svg-O0tQQpKs83eiyBID .section-8 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-8 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-8 path{fill:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-8 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-8{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-8{stroke:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-8{stroke-width:-10;}#mermaid-svg-O0tQQpKs83eiyBID .section-8 line{stroke:hsl(330, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-9 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-9 path,#mermaid-svg-O0tQQpKs83eiyBID .section-9 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-9 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-9 path{fill:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-9 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-9{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-9{stroke:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-9{stroke-width:-13;}#mermaid-svg-O0tQQpKs83eiyBID .section-9 line{stroke:hsl(0, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-10 rect,#mermaid-svg-O0tQQpKs83eiyBID .section-10 path,#mermaid-svg-O0tQQpKs83eiyBID .section-10 circle,#mermaid-svg-O0tQQpKs83eiyBID .section-10 polygon,#mermaid-svg-O0tQQpKs83eiyBID .section-10 path{fill:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-10 text{fill:black;}#mermaid-svg-O0tQQpKs83eiyBID .node-icon-10{font-size:40px;color:black;}#mermaid-svg-O0tQQpKs83eiyBID .section-edge-10{stroke:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .edge-depth-10{stroke-width:-16;}#mermaid-svg-O0tQQpKs83eiyBID .section-10 line{stroke:hsl(30, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-O0tQQpKs83eiyBID .disabled,#mermaid-svg-O0tQQpKs83eiyBID .disabled circle,#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:lightgray;}#mermaid-svg-O0tQQpKs83eiyBID .disabled text{fill:#efefef;}#mermaid-svg-O0tQQpKs83eiyBID .section-root rect,#mermaid-svg-O0tQQpKs83eiyBID .section-root path,#mermaid-svg-O0tQQpKs83eiyBID .section-root circle,#mermaid-svg-O0tQQpKs83eiyBID .section-root polygon{fill:hsl(240, 100%, 46.2745098039%);}#mermaid-svg-O0tQQpKs83eiyBID .section-root text{fill:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .section-root span{color:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .section-2 span{color:#ffffff;}#mermaid-svg-O0tQQpKs83eiyBID .icon-container{height:100%;display:flex;justify-content:center;align-items:center;}#mermaid-svg-O0tQQpKs83eiyBID .edge{fill:none;}#mermaid-svg-O0tQQpKs83eiyBID .mindmap-node-label{dy:1em;alignment-baseline:middle;text-anchor:middle;dominant-baseline:middle;text-align:center;}#mermaid-svg-O0tQQpKs83eiyBID :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} V4.1 Flash
推理系统
要解决的问题
海量用户 + 低价
长上下文 Agent 流量
多厂商加速卡
重复 prefill 浪费
核心概念
E/P/D 分离
Wide-EP 专家并行
分层 KV 缓存
Block KV
SWA KV
SWA Bounded Replay
DSpark 变长验证
统一 Conversation
关键组件
deepseek-recipe
协议 ↔ token
Entrypoint
前缀匹配与路由
负载均衡器
Prefill LB
Decode LB
VE LB
EPLB
KVCache Storage
3FS 约 100 PiB
DeepInfer 引擎
Prefill Service
Decode Service
贯穿示例
编程 Agent 第三轮请求
10 万 token 上下文
带截图 + 工具调用
关键流程
缓存命中路径
SWA 缺失回退
引擎故障转移
3. 阅读路线
整套指南遵循 Why → What → 概念地图 → 贯穿示例 → 组件深潜 → 端到端走查 的顺序。
#mermaid-svg-3WQF5pVNsYreKMnJ{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-3WQF5pVNsYreKMnJ .error-icon{fill:#552222;}#mermaid-svg-3WQF5pVNsYreKMnJ .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-3WQF5pVNsYreKMnJ .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-3WQF5pVNsYreKMnJ .marker{fill:#333333;stroke:#333333;}#mermaid-svg-3WQF5pVNsYreKMnJ .marker.cross{stroke:#333333;}#mermaid-svg-3WQF5pVNsYreKMnJ svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-3WQF5pVNsYreKMnJ p{margin:0;}#mermaid-svg-3WQF5pVNsYreKMnJ .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster-label text{fill:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster-label span{color:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster-label span p{background-color:transparent;}#mermaid-svg-3WQF5pVNsYreKMnJ .label text,#mermaid-svg-3WQF5pVNsYreKMnJ span{fill:#333;color:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ .node rect,#mermaid-svg-3WQF5pVNsYreKMnJ .node circle,#mermaid-svg-3WQF5pVNsYreKMnJ .node ellipse,#mermaid-svg-3WQF5pVNsYreKMnJ .node polygon,#mermaid-svg-3WQF5pVNsYreKMnJ .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-3WQF5pVNsYreKMnJ .rough-node .label text,#mermaid-svg-3WQF5pVNsYreKMnJ .node .label text,#mermaid-svg-3WQF5pVNsYreKMnJ .image-shape .label,#mermaid-svg-3WQF5pVNsYreKMnJ .icon-shape .label{text-anchor:middle;}#mermaid-svg-3WQF5pVNsYreKMnJ .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-3WQF5pVNsYreKMnJ .rough-node .label,#mermaid-svg-3WQF5pVNsYreKMnJ .node .label,#mermaid-svg-3WQF5pVNsYreKMnJ .image-shape .label,#mermaid-svg-3WQF5pVNsYreKMnJ .icon-shape .label{text-align:center;}#mermaid-svg-3WQF5pVNsYreKMnJ .node.clickable{cursor:pointer;}#mermaid-svg-3WQF5pVNsYreKMnJ .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-3WQF5pVNsYreKMnJ .arrowheadPath{fill:#333333;}#mermaid-svg-3WQF5pVNsYreKMnJ .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-3WQF5pVNsYreKMnJ .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-3WQF5pVNsYreKMnJ .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-3WQF5pVNsYreKMnJ .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-3WQF5pVNsYreKMnJ .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-3WQF5pVNsYreKMnJ .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster text{fill:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ .cluster span{color:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-3WQF5pVNsYreKMnJ .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-3WQF5pVNsYreKMnJ rect.text{fill:none;stroke-width:0;}#mermaid-svg-3WQF5pVNsYreKMnJ .icon-shape,#mermaid-svg-3WQF5pVNsYreKMnJ .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-3WQF5pVNsYreKMnJ .icon-shape p,#mermaid-svg-3WQF5pVNsYreKMnJ .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-3WQF5pVNsYreKMnJ .icon-shape .label rect,#mermaid-svg-3WQF5pVNsYreKMnJ .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-3WQF5pVNsYreKMnJ .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-3WQF5pVNsYreKMnJ .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-3WQF5pVNsYreKMnJ :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 01 Why
问题
02 What
定义与边界
03 概念地图
全景
04 贯穿示例
浅层走一遍
05 组件深潜
逐个深入
06 端到端走查
全深度重放
07 下一步
自建与延伸
| 文件 | 内容 |
|---|---|
| 01 为什么 | 推理服务的痛点:成本、长上下文、重复计算、硬件多样性 |
| 02 是什么 | 报告、DeepInfer、deepseek-recipe 各是什么,开源边界在哪 |
| 03 概念地图 | 7 个核心概念 + 7 个关键组件 + 协同总图 |
| 04 贯穿示例 | 贯穿示例:编程 Agent 的第三轮请求 |
| 05.1 deepseek-recipe | 协议转换、防注入编码、分段并行分词、流式解析 |
| 05.2 Entrypoint 与 KV 存储 | 三层 KV 缓存、3FS、命中策略 |
| 05.3 负载均衡器 | Prefill / Decode / VE 负载均衡与 EPLB |
| 05.4 DeepInfer 总览 | DeepInfer 引擎(子系统,内含 4 篇) |
| 05.5 弹性与容错 | 弹性伸缩、冗余配额、EP 故障转移 |
| 06 端到端走查 | 示例端到端全深度重放 |
| 06-walkthrough-interactive.html | 交互式逐跳演示(可切换"SWA 缺失"变体) |
| 07 下一步 | 自建部署清单、开放问题、练习、资料 |
| examples/ | 本地实测的 deepseek-recipe 脚本与输出(渲染 / 防注入 / 流式解析) |
| diagrams/ | 官方架构图与基准图、端到端 draw.io 图 |
4. 缩写对照表
| 缩写 | 英文全称 | 中文 |
|---|---|---|
| API | Application Programming Interface | 应用程序接口 |
| BPE | Byte Pair Encoding | 字节对编码(分词算法) |
| CANN | Compute Architecture for Neural Networks | 华为昇腾异构计算架构 |
| CE | Decode Engine(官方用语,C 指 completion) | 解码引擎实例 |
| CED | Causal Encoder-Decoder | 因果编码器-解码器(V4.1 模型结构) |
| CP | Context Parallelism | 上下文并行 |
| CSA2 | Compressed Sparse Attention 2 | 压缩稀疏注意力第二版 |
| CUDA | Compute Unified Device Architecture | 英伟达通用并行计算架构 |
| DP | Data Parallelism | 数据并行 |
| EBNF | Extended Backus--Naur Form | 扩展巴科斯范式(文法描述) |
| EC | Erasure Coding | 纠删码 |
| EP | Expert Parallelism | 专家并行 |
| EPD | Encoder / Prefill / Decode (disaggregation) | 视觉编码 / 预填充 / 解码分离 |
| EPLB | Expert-Parallel Load Balancer | 专家并行负载均衡 |
| FIFO | First In, First Out | 先进先出 |
| GC | Garbage Collection | 垃圾回收 |
| GDR | GPUDirect RDMA | GPU 直连远程内存访问 |
| H2D | Host to Device | 主机内存到显存 |
| HBM | High Bandwidth Memory | 高带宽显存 |
| KMP | Knuth--Morris--Pratt | KMP 字符串匹配算法 |
| KV | Key-Value (cache) | 注意力键值缓存 |
| LB | Load Balancer | 负载均衡器 |
| LLM | Large Language Model | 大语言模型 |
| LRU | Least Recently Used | 最近最少使用 |
| MoE | Mixture of Experts | 混合专家模型 |
| MTP | Multi-Token Prediction | 多 token 预测(草稿头) |
| MXFP4 / MXFP8 | Microscaling FP4 / FP8 | 微缩放 4 位 / 8 位浮点格式 |
| NPU | Neural Processing Unit | 神经网络处理器 |
| PCIe | Peripheral Component Interconnect Express | 高速外设互连总线 |
| PE | Prefill Engine | 预填充引擎实例 |
| RDMA | Remote Direct Memory Access | 远程直接内存访问 |
| SM | Streaming Multiprocessor | 流式多处理器(GPU 计算单元;SM100 指 Blackwell 架构) |
| SO | Structured Output | 结构化输出 |
| SSD | Solid State Drive | 固态硬盘 |
| SWA | Sliding Window Attention | 滑动窗口注意力 |
| TBT | Time Between Tokens | token 间隔时延 |
| TP | Tensor Parallelism | 张量并行 |
| TPS | Tokens Per Second | 每秒 token 数 |
| TTFT | Time To First Token | 首 token 时延 |
| TTL | Time To Live | 存活时间 |
| VE | Vision Encoder | 视觉编码器 |
| VLAN | Virtual Local Area Network | 虚拟局域网 |
| YOCO | You Only Cache Once | "只缓存一次"跨层 KV 复用结构 |
| 3FS | Fire-Flyer File System | DeepSeek 开源的分布式文件系统 |
| DSML | (官方未展开全称)DeepSeek 工具调用标记语法 | 模型输出工具调用时使用的标签格式 |
5. 资料来源与可信度说明
| 来源 | 用途 | 可信度 |
|---|---|---|
DeepInfer doc/inference_deployment_qa.md |
官方 Q&A,明确声明"基于知乎技术报告",本指南绝大多数数字出自此处 | 官方一手 |
DeepInfer README 与 doc/images/*.png |
功能列表、环境变量、硬件、基准图 | 官方一手 |
| deepseek-recipe README / docs | 包结构、API 示例、分词与流式文档 | 官方一手 |
| arXiv 2609.19969 摘要 | 模型参数量、CED 结构、KV 字节数 | 官方论文摘要 |
| (https://www.sysgeek.cn/deepseek-deepinfer-open-source/) | 首 token 约 0.4 秒、230 token/秒等媒体转述数字 | 二手,文中标注 |