jshn9515 · Oct 3, 2026 · Oct 3, 2026 · Oct 3, 2026 · Oct 3, 2026
diff --git a/_quarto-html.yml b/_quarto-html.yml
@@ -108,8 +108,8 @@ website:
               contents: "zh/ch12-llm-training-engineering/*.qmd"
         - section: "Part 7: LLM Serving：vLLM 与 SGLang"
           contents:
-            - section: "Chapter 19: SGLang：从请求调度到 Cache-Aware Serving"
-              contents: "zh/ch19-sglang-serving/*.qmd"
+            - section: "Chapter 18: vLLM：从 LLM 推理到高性能推理框架"
+              contents: "zh/ch18-vllm-introduction/*.qmd"
     - id: en
       style: floating
       collapse-level: 1

diff --git a/_quarto-typst-zh.yml b/_quarto-typst-zh.yml
@@ -148,6 +148,6 @@ book:
         - "zh/ch12-llm-training-engineering/ch12.9-distributed-training.qmd"
         - "zh/ch12-llm-training-engineering/ch12.10-large-model-checkpoint.qmd"
         - "zh/ch12-llm-training-engineering/ch12.11-exercises.qmd"
-    - part: "Chapter 19: SGLang：从请求调度到 Cache-Aware Serving"
+    - part: "Chapter 18: vLLM：从 LLM 推理到高性能推理框架"
       chapters:
-        - "zh/ch19-sglang-serving/ch19.1-from-vllm-to-sglang.qmd"
+        - "zh/ch18-vllm-introduction/ch18.1-vllm-overview.qmd"
diff --git a/refs-article.bib b/refs-article.bib
@@ -197,6 +197,15 @@ @article{kingma2022VAE
   url           = {https://arxiv.org/abs/1312.6114},
   year          = {2022}
 }
+@article{kwon2023vLLM,
+  archiveprefix = {arXiv},
+  author        = {Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
+  eprint        = {2309.06180},
+  primaryclass  = {cs.LG},
+  title         = {Efficient Memory Management for Large Language Model Serving with PagedAttention},
+  url           = {https://arxiv.org/abs/2309.06180},
+  year          = {2023}
+}
 @article{li2022BLIP,
   archiveprefix = {arXiv},
   author        = {Junnan Li and Dongxu Li and Caiming Xiong and Steven Hoi},

diff --git a/zh/README.md b/zh/README.md
@@ -112,6 +112,6 @@
 - 12.10 大模型 Checkpoint：模型、优化器与分布式状态如何恢复
 - 12.11 Chapter 12 练习题：LLM 训练工程
 
-## Chapter 19: SGLang：从请求调度到 Cache-Aware Serving
+## Chapter 18: vLLM：从 LLM 推理到高性能推理框架
 
-- 19.1 From vLLM to SGLang：相同的 Serving 问题，不同的设计
+- 18.1 从 generate() 到 vLLM：为什么推理需要一个 Engine
diff --git a/zh/ch18-vllm-introduction/ch18.1-vllm-overview.qmd b/zh/ch18-vllm-introduction/ch18.1-vllm-overview.qmd
diff --git a/zh/ch18-vllm-introduction/figures/ch18.1-req-lifecycle.mmd b/zh/ch18-vllm-introduction/figures/ch18.1-req-lifecycle.mmd
@@ -0,0 +1,46 @@
+---
+config:
+  startOnLoad: false
+  theme: base
+  htmlLabels: false
+  fontFamily: Arial, Microsoft YaHei, Noto Serif CJK SC, sans-serif
+  themeVariables:
+    fontSize: 16px
+    primaryColor: "#e9f2fc"
+    primaryBorderColor: "#296bb7"
+    primaryTextColor: "#333333"
+    lineColor: "#0b0b0b"
+
+  themeCSS: |
+    .node text,
+    .nodeLabel,
+    .label text {
+      font-weight: 400;
+      font-style: normal;
+    }
+
+    .node rect,
+    .flowchart-link {
+      stroke-width: 1px;
+    }
+---
+
+sequenceDiagram
+    participant A as API Server
+    participant E as EngineCore
+    participant S as Scheduler
+    participant K as KVCacheManager
+    participant G as GPU Worker
+
+    A->>E: 发送 Token IDs (IPC)
+    E->>S: 加入请求
+    E->>S: schedule()
+    S->>K: 分配/检查 KV Cache
+    K-->>S: 缓存分配结果
+    S-->>E: SchedulerOutput
+    E->>G: 执行模型 (via Executor)
+    G-->>E: ModelRunnerOutput
+    E->>S: update_from_output()
+    E-->>A: 输出 token (IPC)
+    Note over E,G: 持续调度和执行直到请求完成
+    A-->>A: Detokenization
diff --git a/zh/ch18-vllm-introduction/figures/ch18.1-req-lifecycle.svg b/zh/ch18-vllm-introduction/figures/ch18.1-req-lifecycle.svg
diff --git a/zh/ch18-vllm-introduction/figures/ch18.1-vllm-architecture.svg b/zh/ch18-vllm-introduction/figures/ch18.1-vllm-architecture.svg
diff --git a/zh/ch18-vllm-introduction/figures/ch18.1-vllm-architecture.xml b/zh/ch18-vllm-introduction/figures/ch18.1-vllm-architecture.xml
@@ -0,0 +1,111 @@
+<mxfile host="Electron">
+  <diagram name="vLLM V1 Architecture" id="U81PoWHJHGe-g25c4hQW">
+    <mxGraphModel dx="1563" dy="1128" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="827" pageHeight="1169" math="0" shadow="0">
+      <root>
+        <mxCell id="0" />
+        <mxCell id="1" parent="0" />
+        <mxCell id="hyPA-xEpt3381CLIv8I8-1" parent="1" style="rounded=1;whiteSpace=wrap;html=1;container=0;fillColor=#F0F6FF;strokeColor=#76A7EC;strokeWidth=2;arcSize=7;align=center;verticalAlign=top;spacingTop=14;fontSize=26;fontStyle=1;fontColor=#22375A;shadow=1;convertToSvg=1;svgWhiteSpace=wrap;" value="Offline" vertex="1">
+          <mxGeometry height="240" width="390" x="350" y="20" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-2" parent="1" style="rounded=1;whiteSpace=wrap;html=1;container=0;fillColor=#F0FAF4;strokeColor=#76B995;strokeWidth=2;arcSize=7;align=center;verticalAlign=top;spacingTop=14;fontSize=26;fontStyle=1;fontColor=#24533B;shadow=1;convertToSvg=1;svgWhiteSpace=wrap;" value="Online · API Server Process" vertex="1">
+          <mxGeometry height="240" width="390" x="770" y="20" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-3" parent="1" style="rounded=1;whiteSpace=wrap;html=1;container=0;fillColor=#FFF9EF;strokeColor=#E8B85B;strokeWidth=2;arcSize=6;align=center;verticalAlign=top;spacingTop=14;fontSize=18;fontStyle=1;fontColor=#694A1B;shadow=1;convertToSvg=1;svgWhiteSpace=wrap;" value="&lt;font style=&quot;font-size: 26px;&quot;&gt;EngineCore Process&lt;/font&gt;" vertex="1">
+          <mxGeometry height="380" width="635" x="442.5" y="410" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-4" parent="1" style="rounded=1;whiteSpace=wrap;html=1;container=0;fillColor=#FFF2F4;strokeColor=#E590A0;strokeWidth=2;arcSize=7;align=center;verticalAlign=top;spacingTop=14;fontSize=18;fontStyle=1;fontColor=#80394B;shadow=1;convertToSvg=1;svgWhiteSpace=wrap;" value="&lt;font style=&quot;font-size: 26px;&quot;&gt;GPU Worker Process&lt;/font&gt;" vertex="1">
+          <mxGeometry height="256" width="315" x="40" y="534" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-5" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#E3EEFF;strokeColor=#6498DF;strokeWidth=1.8;fontColor=#1B2E4B;fontSize=22;shadow=1;arcSize=9;fontFamily=Helvetica;convertToSvg=1;svgWhiteSpace=wrap;" value="LLM.chat / LLM.generate" vertex="1">
+          <mxGeometry height="54" width="285" x="402" y="85" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-6" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#E3EEFF;strokeColor=#6498DF;strokeWidth=1.8;fontColor=#1B2E4B;fontSize=22;shadow=1;arcSize=9;fontFamily=Helvetica;convertToSvg=1;svgWhiteSpace=wrap;" value="LLMEngine" vertex="1">
+          <mxGeometry height="52" width="135" x="477" y="175" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-7" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#E2F6E9;strokeColor=#6AAF85;strokeWidth=1.8;fontColor=#1B4931;fontSize=22;shadow=1;arcSize=9;fontFamily=Helvetica;convertToSvg=1;svgWhiteSpace=wrap;" value="HTTP Request" vertex="1">
+          <mxGeometry height="54" width="195" x="867" y="85" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-8" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#E2F6E9;strokeColor=#6AAF85;strokeWidth=1.8;fontColor=#1B4931;fontSize=22;shadow=1;arcSize=9;fontFamily=Helvetica;convertToSvg=1;svgWhiteSpace=wrap;" value="API Server / AsyncLLM" vertex="1">
+          <mxGeometry height="52" width="245" x="842" y="175" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-9" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#EFE8FF;strokeColor=#9280D5;strokeWidth=1.8;fontColor=#392D65;fontSize=24;fontStyle=1;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="EngineCoreClient" vertex="1">
+          <mxGeometry height="58" width="240" x="640" y="310" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-10" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFEAC8;strokeColor=#E3AB47;strokeWidth=1.8;fontColor=#503913;fontSize=22;fontStyle=1;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="EngineCore" vertex="1">
+          <mxGeometry height="54" width="201" x="659" y="480" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-11" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFEAC8;strokeColor=#E3AB47;strokeWidth=1.8;fontColor=#503913;fontSize=22;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="Model Executor" vertex="1">
+          <mxGeometry height="56" width="190" x="493" y="602" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-12" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFEAC8;strokeColor=#E3AB47;strokeWidth=1.8;fontColor=#503913;fontSize=22;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="Scheduler" vertex="1">
+          <mxGeometry height="56" width="165" x="844" y="602" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-13" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFEAC8;strokeColor=#E3AB47;strokeWidth=1.8;fontColor=#503913;fontSize=22;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="KVCacheManager" vertex="1">
+          <mxGeometry height="56" width="201" x="826" y="706" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-14" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFE1E8;strokeColor=#DF8396;strokeWidth=1.8;fontColor=#713243;fontSize=22;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="GPU Worker" vertex="1">
+          <mxGeometry height="55" width="165" x="115" y="612" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-15" parent="1" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#FFE1E8;strokeColor=#DF8396;strokeWidth=1.8;fontColor=#713243;fontSize=22;shadow=1;arcSize=9;convertToSvg=1;svgWhiteSpace=wrap;" value="Model Runner" vertex="1">
+          <mxGeometry height="55" width="165" x="115" y="706.5" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-16" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-5" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=0.5;entryY=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-6">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-17" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-7" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=0.5;entryY=0;fontSize=12;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-8">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-18" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-6" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;fontSize=12;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-9">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="544.4300000000001" y="339" />
+            </Array>
+            <mxPoint x="620" y="330" as="targetPoint" />
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-19" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-8" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=1;entryY=0.5;entryDx=0;entryDy=0;fontSize=12;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-9">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="964.43" y="339" />
+            </Array>
+            <mxPoint x="1090" y="320" as="targetPoint" />
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-20" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-10" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;entryX=0.5;entryY=0;exitX=0.5;exitY=1;exitDx=0;exitDy=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-11">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points" />
+            <mxPoint x="730" y="534" as="sourcePoint" />
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-21" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-10" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;entryX=0.5;entryY=0;exitX=0.5;exitY=1;exitDx=0;exitDy=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-12">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points" />
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-22" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-12" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=0.5;entryY=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-13">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-23" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-11" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-4">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="588" y="710" />
+              <mxPoint x="388" y="710" />
+              <mxPoint x="388" y="470" />
+              <mxPoint x="197.53" y="470" />
+            </Array>
+            <mxPoint x="190" y="490" as="targetPoint" />
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-24" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-14" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;endArrow=block;endFill=1;strokeColor=#334155;strokeWidth=2;exitX=0.5;exitY=1;entryX=0.5;entryY=0;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-15">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="hyPA-xEpt3381CLIv8I8-25" edge="1" parent="1" source="hyPA-xEpt3381CLIv8I8-9" style="endArrow=block;html=1;rounded=1;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;strokeWidth=2;strokeColor=#334155;align=center;verticalAlign=middle;fontFamily=Helvetica;fontSize=12;fontColor=default;labelBackgroundColor=default;endFill=1;convertToSvg=1;" target="hyPA-xEpt3381CLIv8I8-3" value="">
+          <mxGeometry height="50" relative="1" width="50" as="geometry">
+            <mxPoint x="840" y="500" as="sourcePoint" />
+            <mxPoint x="890" y="450" as="targetPoint" />
+          </mxGeometry>
+        </mxCell>
+      </root>
+    </mxGraphModel>
+  </diagram>
+</mxfile>
diff --git a/zh/index-parts/part07.qmd b/zh/index-parts/part07.qmd
@@ -8,7 +8,7 @@ comments: false
 
 listing:
   contents:
-    - "../ch19-sglang-serving/*.qmd"
+    - "../ch18-vllm-introduction/*.qmd"
   sort: title asc
   type: table
   fields: [title, author, date]