[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-47":104},{"enabled":4,"title":5,"statusText":5,"buttonText":5,"campSlug":5,"stage":6,"classroomId":7,"hasClassroom":4},false,"",0,null,{"total":9,"list":10,"pageNum":44,"pageSize":45,"size":9,"startRow":44,"endRow":9,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":48,"navigateFirstPage":44,"navigateLastPage":44},3,[11,25,35],{"slug":12,"title":13,"logo":7,"liveUrl":14,"prize":15,"sponsor":16,"description":17,"organizer":18,"startTime":19,"endTime":20,"registrationDeadline":21,"status":22,"dashboardVisibility":23,"createdAt":24},"spring2026","2026春季启元人工智能大赛","https:\u002F\u002Fwww.bilibili.com\u002Fblackboard\u002Flive\u002Flive-activity-player.html?cid=1822703816&quality=0","200,000","麒麟软件有限公司","“2026春季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。\n各参赛选手需在规定的比赛周期内，依据赛事规则选定赛题并完成相应挑战，优胜者将共同角逐总额20万元的赛事奖金池，并且还将有机会获得实习和知名企业优质岗位内推资格。","启元实验室","2026-04-19 16:00:00","2026-07-13 16:00:00","2026-06-16 16:00:00","ongoing","private","2026-04-16 08:00:00",{"slug":26,"title":27,"logo":7,"liveUrl":14,"prize":28,"sponsor":7,"description":29,"organizer":18,"startTime":30,"endTime":31,"registrationDeadline":32,"status":33,"dashboardVisibility":23,"createdAt":34},"autumn2025","2025秋季启元人工智能大赛","150,000","“2025秋季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。","2025-11-08 08:00:00","2026-01-13 08:00:00","2025-12-30 08:00:00","ended","2025-11-03 08:00:00",{"slug":36,"title":37,"logo":7,"liveUrl":14,"prize":38,"sponsor":7,"description":39,"organizer":18,"startTime":40,"endTime":41,"registrationDeadline":42,"status":33,"dashboardVisibility":23,"createdAt":43},"summer2025","启元人工智能大赛(系统方向)","80,000","启元人工智能大赛由启元实验室主办，其中系统赛道基于启元实验室九源软件栈的相关开源项目，通过拉动开源社区力量，一方面促进国产智能计算生态的发展，另一方面也为人工智能系统方向的学生及从业者提供一个实践平台。","2025-07-08 08:00:00","2025-08-25 08:00:00","2025-07-22 08:00:00","2025-06-22 08:00:00",1,5,true,8,[44],{"total":45,"list":50,"pageNum":44,"pageSize":45,"size":45,"startRow":44,"endRow":45,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":103,"navigateFirstPage":44,"navigateLastPage":44},[51,62,73,84,94],{"id":52,"slug":53,"title":54,"logo":55,"publicLiveEmbedUrl":14,"description":56,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":16,"startTime":58,"endTime":59,"status":22,"dashboardVisibility":60,"createdAt":58,"updatedAt":61},"019eeeff-6b00-7d55-a629-efafded3e232","summer2026","2026年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2026-06\u002F019eef7f-7816-7828-9309-79c2c61dd568.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fLKIdTmv5ztTF5lNyW4vFO%2FwiU4%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。","...","2026-06-22 08:00:00","2026-10-19 08:00:00","public","2026-10-12 08:00:00",{"id":63,"slug":64,"title":65,"logo":66,"publicLiveEmbedUrl":7,"description":67,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":68,"startTime":69,"endTime":70,"status":33,"dashboardVisibility":23,"createdAt":71,"updatedAt":72},"019af3f3-e06a-7245-aa8d-37dc99a69c73","winter2025","2025年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-12\u002F019afc69-b7ca-7805-91fb-605429671c01.png?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IfcKyBkOe6Ug2cb8ODO9I1IUNkA%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。 \n","新华三技术有限公司","2025-12-25 08:00:00","2026-03-27 08:00:00","2025-12-06 08:00:00","2026-10-22 20:24:30",{"id":74,"slug":36,"title":75,"logo":76,"publicLiveEmbedUrl":7,"description":77,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":78,"startTime":79,"endTime":80,"status":81,"dashboardVisibility":23,"createdAt":82,"updatedAt":83},"01980342-208c-7803-b3f3-19e1d31960ba","2025年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-07\u002F0198045e-28fe-78f0-95fe-23835ce58395.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ev4DWdZSXJ0t1T4u2JL5qybAazs%3D","InfiniTensor大模型与人工智能系统训练营由启元实验室及清华大学 AI 系统方向团队发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营联合开源项目开发者、学术界大牛、工业界资深工程师参与授课与项目指导。课程采用「线上直播授课 + 在线答疑 + 项目实践辅导 」的方式进行，确保每位同学都能获得充分成长。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","北京积算科技有限公司","2025-07-14 08:00:00","2025-09-30 08:00:00","upcoming","2025-07-13 08:00:00","2026-10-22 20:24:58",{"id":85,"slug":86,"title":87,"logo":88,"publicLiveEmbedUrl":7,"description":89,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":90,"endTime":91,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":93},"01980cf9-a7d8-7796-ae86-0bbe05334add","winter2024","2024年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-ce23-7624-a287-9f07dd0d85f6.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。报名学员请阅读“导学”课程页的训练营介绍。","2025-01-05 08:00:00","2025-03-22 08:00:00","2025-07-15 08:00:00","2026-10-22 20:24:39",{"id":95,"slug":96,"title":97,"logo":98,"publicLiveEmbedUrl":7,"description":99,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":100,"endTime":101,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":102},"01980c20-363a-7dc3-8f5a-58a68e08dcc4","summer2024","2024年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-fbfa-788a-afcd-17fa9be74ff9.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区和高校，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过课程、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","2024-07-20 08:00:00","2024-09-29 08:00:00","2026-10-22 20:24:49",[44],{"currentItem":105,"processedData":125},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":117,"canonicalPath":118,"id":119,"contentType":120,"prevArticle":121},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_fm.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E24n%2FXmx%2BroiIyrXbPqr6wFxJ68%3D","InfiniTensor,冬季训练营,大模型服务系统,服务架构,请求调度","聚焦大模型服务系统设计，涵盖服务架构、请求调度、稳定性与可扩展性建设。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营推理引擎第五课《大模型服务系统》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z0uKWOaV3A1%2F0KGTIvz4s%2FbVFcI%3D)\r\n\r\n\r\n\r\n\r\n2026 年 1 月 29 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎系列第五课《大模型服务系统》开讲。\r\n\r\n本节课聚焦大模型推理服务系统架构，讲解单用户推理服务与多用户推理服务区别。\r\n\r\n## 大模型推理服务基本功能\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bOadBOPU2x9BrX%2B5gfEvGBR07ow%3D)\r\n\r\n### 1.用户交互功能\r\n- 用户输入与模型推理\r\n- 流式输出与中断推理\r\n- 多轮对话与历史问题修改\r\n- 历史会话记录与管理\r\n### 2.会话管理\r\n- 会话管理的核心之一是 KV Cache 管理，但同时还包括对话历史、上下文裁剪策略及会话生命周期控制等内容\r\n- KV Cache 的内存分配与最大上下文长度\r\n- 对话总长度超过最大上下文长度的处理方法\r\n  - 保留最近 n 轮对话\r\n  - 对更早的对话进行总结\r\n  - 对用户所有历史对话进行归纳\r\n\r\n\r\n## 单用户、多会话实现与优化\r\n\r\n### 1.简单实现方法：\r\n- 每个会话绑定一个 KV cache\r\n- 切换会话时切换 KV cache\r\n- 优点：会话 KV cache 命中率高\r\n- 缺点：KV cache 不能无限分配\r\n\r\n### 2.优化：KV cache 池\r\n- 固定数量 KV cache\r\n- 前缀匹配方式分配\r\n- 缺点：总会话数受限，内存利用率低\r\n\r\n### 3.优化：Paged Attention\u003Cbr>（vLLM 核心创新）\r\n- 使用“内存页”思想\r\n- 动态分配内存块\r\n- 内存共享与引用计数\r\n- 在显存约束下支持大规模并发会话\r\n- 复用与内存回收机制\r\n\r\n## 多用户服务实现\r\n\r\n### 1.多用户与单用户区别：\r\n- 多用户服务中，多个用户通过网络接口同时向推理服务发送请求，服务端需要处理并发与调度问题\r\n- 任意数量的请求随时可能出现，服务需要第一时间响应\r\n\r\n### 2.标准 API 介绍\r\n- 业界广泛采用并兼容的 OpenAI 风格 API 接口\r\n- JSON 格式数据传输\r\n- 流式与非流式接口\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qhx76cagW7E8TONC1TdCHSfa%2BRA%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XbcNZzPzAY29uqS1%2F7hDPFAwavQ%3D)\r\n\r\n\r\n### 3.服务层功能\r\n- 接收用户请求\r\n- 请求队列管理\r\n- 身份认证与网络分流\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HiVivaFmn8BFmcO%2BcRBPnT7SUfw%3D)\r\n\r\n\r\n### 4.调度器功能\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xQ17ISbe9V17V48imzHcX89lA34%3D)\r\n\r\n#### （1） 服务多请求\r\n#### a. 以迭代为单位形成 batch\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=q7%2BavMaZ9sZLOJ0znAC796ZrqWQ%3D)\r\n\r\n\r\n#### b. 跨请求批量化计算：\r\n- 可批量化计算部分（MLP 层、RMS norm 等）\r\n- 批量化难度较高的部分（Attention 相关的）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OGjfobHMQdge6iCfq3qalxmO75E%3D)\r\n\r\n#### c. Orca 策略：连续批次\r\n- 请求池管理\r\n- 迭代级批次形成与计算\r\n- 混合 PD 请求处理\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=k4S3uev6ETyJdodi%2Fu04nOK24vg%3D)\r\n\r\n\r\n#### （2） PD 分离优化\r\n- Prefill 与 Decode 阶段分离\r\n- 不同硬件资源分配\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MwHtHOMaclbm5b2ketpnS4ZB0jA%3D)\r\n\r\n\r\n#### （3） Chunk Prefill 与 Decode-Only Batch\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=K2lodhrNNX9I2sUdW5OqGZ93nVM%3D)\r\n\r\n### 模型层\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGxxT1undEBpali5U1Rq5i9Lmwo%3D)\r\n\r\n\r\n\r\n## 服务系统架构设计示例\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1RQEbiOTzRRSI8cP9euDE6C%2FHuA%3D)\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n![2025冬季训练营课程表](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6oUxPj5AAym697VnYMYJWkGBgLo%3D)\r\n\r\n### 💡 学习贴士\r\n更多细节及完整课程内容，请观看直播或查看课程回放。\r\n\r\n链接：https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\r\n\r\n### 📚 训练营报名持续进行中\r\n扫码报名与加入群聊，或点击 **「阅读原文」** 直达官网报名\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sJcS84TrEl4MrM0JkfFwK7feHzA%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-46","study-46","InfiniTensor 冬季训练营训练系统第四课《训练框架编程基础》",62,"\u002Fstudy-material\u002Fstudy-47","study-47","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-48","study-48","InfiniTensor 冬季训练营 Triton&九齿第七课《九齿三重：天通》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z0uKWOaV3A1%2F0KGTIvz4s%2FbVFcI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 29 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎系列第五课《大模型服务系统》开讲。\u003C\u002Fp>\n\u003Cp>本节课聚焦大模型推理服务系统架构，讲解单用户推理服务与多用户推理服务区别。\u003C\u002Fp>\n\u003Ch2 id=\"大模型推理服务基本功能\">\u003Ca class=\"anchor\" href=\"#大模型推理服务基本功能\" aria-label=\"跳转到标题：大模型推理服务基本功能\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>大模型推理服务基本功能\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bOadBOPU2x9BrX%2B5gfEvGBR07ow%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"1-用户交互功能\">\u003Ca class=\"anchor\" href=\"#1-用户交互功能\" aria-label=\"跳转到标题：1.用户交互功能\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.用户交互功能\u003C\u002Fh3>\u003Cul>\n\u003Cli>用户输入与模型推理\u003C\u002Fli>\n\u003Cli>流式输出与中断推理\u003C\u002Fli>\n\u003Cli>多轮对话与历史问题修改\u003C\u002Fli>\n\u003Cli>历史会话记录与管理\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"2-会话管理\">\u003Ca class=\"anchor\" href=\"#2-会话管理\" aria-label=\"跳转到标题：2.会话管理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.会话管理\u003C\u002Fh3>\u003Cul>\n\u003Cli>会话管理的核心之一是 KV Cache 管理，但同时还包括对话历史、上下文裁剪策略及会话生命周期控制等内容\u003C\u002Fli>\n\u003Cli>KV Cache 的内存分配与最大上下文长度\u003C\u002Fli>\n\u003Cli>对话总长度超过最大上下文长度的处理方法\u003Cul>\n\u003Cli>保留最近 n 轮对话\u003C\u002Fli>\n\u003Cli>对更早的对话进行总结\u003C\u002Fli>\n\u003Cli>对用户所有历史对话进行归纳\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"单用户-多会话实现与优化\">\u003Ca class=\"anchor\" href=\"#单用户-多会话实现与优化\" aria-label=\"跳转到标题：单用户、多会话实现与优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>单用户、多会话实现与优化\u003C\u002Fh2>\u003Ch3 id=\"1-简单实现方法\">\u003Ca class=\"anchor\" href=\"#1-简单实现方法\" aria-label=\"跳转到标题：1.简单实现方法：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.简单实现方法：\u003C\u002Fh3>\u003Cul>\n\u003Cli>每个会话绑定一个 KV cache\u003C\u002Fli>\n\u003Cli>切换会话时切换 KV cache\u003C\u002Fli>\n\u003Cli>优点：会话 KV cache 命中率高\u003C\u002Fli>\n\u003Cli>缺点：KV cache 不能无限分配\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"2-优化-kv-cache-池\">\u003Ca class=\"anchor\" href=\"#2-优化-kv-cache-池\" aria-label=\"跳转到标题：2.优化：KV cache 池\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.优化：KV cache 池\u003C\u002Fh3>\u003Cul>\n\u003Cli>固定数量 KV cache\u003C\u002Fli>\n\u003Cli>前缀匹配方式分配\u003C\u002Fli>\n\u003Cli>缺点：总会话数受限，内存利用率低\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"3-优化-paged-attention-vllm-核心创新\">\u003Ca class=\"anchor\" href=\"#3-优化-paged-attention-vllm-核心创新\" aria-label=\"跳转到标题：3.优化：Paged Attention（vLLM 核心创新）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.优化：Paged Attention\u003Cbr>（vLLM 核心创新）\u003C\u002Fh3>\u003Cul>\n\u003Cli>使用“内存页”思想\u003C\u002Fli>\n\u003Cli>动态分配内存块\u003C\u002Fli>\n\u003Cli>内存共享与引用计数\u003C\u002Fli>\n\u003Cli>在显存约束下支持大规模并发会话\u003C\u002Fli>\n\u003Cli>复用与内存回收机制\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"多用户服务实现\">\u003Ca class=\"anchor\" href=\"#多用户服务实现\" aria-label=\"跳转到标题：多用户服务实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>多用户服务实现\u003C\u002Fh2>\u003Ch3 id=\"1-多用户与单用户区别\">\u003Ca class=\"anchor\" href=\"#1-多用户与单用户区别\" aria-label=\"跳转到标题：1.多用户与单用户区别：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.多用户与单用户区别：\u003C\u002Fh3>\u003Cul>\n\u003Cli>多用户服务中，多个用户通过网络接口同时向推理服务发送请求，服务端需要处理并发与调度问题\u003C\u002Fli>\n\u003Cli>任意数量的请求随时可能出现，服务需要第一时间响应\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"2-标准-api-介绍\">\u003Ca class=\"anchor\" href=\"#2-标准-api-介绍\" aria-label=\"跳转到标题：2.标准 API 介绍\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.标准 API 介绍\u003C\u002Fh3>\u003Cul>\n\u003Cli>业界广泛采用并兼容的 OpenAI 风格 API 接口\u003C\u002Fli>\n\u003Cli>JSON 格式数据传输\u003C\u002Fli>\n\u003Cli>流式与非流式接口\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qhx76cagW7E8TONC1TdCHSfa%2BRA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XbcNZzPzAY29uqS1%2F7hDPFAwavQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"3-服务层功能\">\u003Ca class=\"anchor\" href=\"#3-服务层功能\" aria-label=\"跳转到标题：3.服务层功能\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.服务层功能\u003C\u002Fh3>\u003Cul>\n\u003Cli>接收用户请求\u003C\u002Fli>\n\u003Cli>请求队列管理\u003C\u002Fli>\n\u003Cli>身份认证与网络分流\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HiVivaFmn8BFmcO%2BcRBPnT7SUfw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"4-调度器功能\">\u003Ca class=\"anchor\" href=\"#4-调度器功能\" aria-label=\"跳转到标题：4.调度器功能\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4.调度器功能\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xQ17ISbe9V17V48imzHcX89lA34%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-服务多请求\">\u003Ca class=\"anchor\" href=\"#1-服务多请求\" aria-label=\"跳转到标题：（1） 服务多请求\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（1） 服务多请求\u003C\u002Fh4>\u003Ch4 id=\"a-以迭代为单位形成-batch\">\u003Ca class=\"anchor\" href=\"#a-以迭代为单位形成-batch\" aria-label=\"跳转到标题：a. 以迭代为单位形成 batch\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>a. 以迭代为单位形成 batch\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=q7%2BavMaZ9sZLOJ0znAC796ZrqWQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"b-跨请求批量化计算\">\u003Ca class=\"anchor\" href=\"#b-跨请求批量化计算\" aria-label=\"跳转到标题：b. 跨请求批量化计算：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>b. 跨请求批量化计算：\u003C\u002Fh4>\u003Cul>\n\u003Cli>可批量化计算部分（MLP 层、RMS norm 等）\u003C\u002Fli>\n\u003Cli>批量化难度较高的部分（Attention 相关的）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OGjfobHMQdge6iCfq3qalxmO75E%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"c-orca-策略-连续批次\">\u003Ca class=\"anchor\" href=\"#c-orca-策略-连续批次\" aria-label=\"跳转到标题：c. Orca 策略：连续批次\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>c. Orca 策略：连续批次\u003C\u002Fh4>\u003Cul>\n\u003Cli>请求池管理\u003C\u002Fli>\n\u003Cli>迭代级批次形成与计算\u003C\u002Fli>\n\u003Cli>混合 PD 请求处理\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=k4S3uev6ETyJdodi%2Fu04nOK24vg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-pd-分离优化\">\u003Ca class=\"anchor\" href=\"#2-pd-分离优化\" aria-label=\"跳转到标题：（2） PD 分离优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（2） PD 分离优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>Prefill 与 Decode 阶段分离\u003C\u002Fli>\n\u003Cli>不同硬件资源分配\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MwHtHOMaclbm5b2ketpnS4ZB0jA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-chunk-prefill-与-decode-only-batch\">\u003Ca class=\"anchor\" href=\"#3-chunk-prefill-与-decode-only-batch\" aria-label=\"跳转到标题：（3） Chunk Prefill 与 Decode-Only Batch\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（3） Chunk Prefill 与 Decode-Only Batch\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=K2lodhrNNX9I2sUdW5OqGZ93nVM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"模型层\">\u003Ca class=\"anchor\" href=\"#模型层\" aria-label=\"跳转到标题：模型层\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型层\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kGxxT1undEBpali5U1Rq5i9Lmwo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"服务系统架构设计示例\">\u003Ca class=\"anchor\" href=\"#服务系统架构设计示例\" aria-label=\"跳转到标题：服务系统架构设计示例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>服务系统架构设计示例\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_47_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1RQEbiOTzRRSI8cP9euDE6C%2FHuA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6oUxPj5AAym697VnYMYJWkGBgLo%3D\" alt=\"2025冬季训练营课程表\">\u003C\u002Fp>\n\u003Ch3 id=\"学习贴士\">\u003Ca class=\"anchor\" href=\"#学习贴士\" aria-label=\"跳转到标题：💡 学习贴士\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>💡 学习贴士\u003C\u002Fh3>\u003Cp>更多细节及完整课程内容，请观看直播或查看课程回放。\u003C\u002Fp>\n\u003Cp>链接：\u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\">https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch3 id=\"训练营报名持续进行中\">\u003Ca class=\"anchor\" href=\"#训练营报名持续进行中\" aria-label=\"跳转到标题：📚 训练营报名持续进行中\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📚 训练营报名持续进行中\u003C\u002Fh3>\u003Cp>扫码报名与加入群聊，或点击 \u003Cstrong>「阅读原文」\u003C\u002Fstrong> 直达官网报名\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sJcS84TrEl4MrM0JkfFwK7feHzA%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,133,136,139,142,145,148,150,153,156,159,162,166,169,172,175,178,181,183,185,188],{"id":129,"text":129,"level":110},"大模型推理服务基本功能",{"id":131,"text":132,"level":9},"1-用户交互功能","1.用户交互功能",{"id":134,"text":135,"level":9},"2-会话管理","2.会话管理",{"id":137,"text":138,"level":110},"单用户-多会话实现与优化","单用户、多会话实现与优化",{"id":140,"text":141,"level":9},"1-简单实现方法","1.简单实现方法：",{"id":143,"text":144,"level":9},"2-优化-kv-cache-池","2.优化：KV cache 池",{"id":146,"text":147,"level":9},"3-优化-paged-attention-vllm-核心创新","3.优化：Paged Attention（vLLM 核心创新）",{"id":149,"text":149,"level":110},"多用户服务实现",{"id":151,"text":152,"level":9},"1-多用户与单用户区别","1.多用户与单用户区别：",{"id":154,"text":155,"level":9},"2-标准-api-介绍","2.标准 API 介绍",{"id":157,"text":158,"level":9},"3-服务层功能","3.服务层功能",{"id":160,"text":161,"level":9},"4-调度器功能","4.调度器功能",{"id":163,"text":164,"level":165},"1-服务多请求","（1） 服务多请求",4,{"id":167,"text":168,"level":165},"a-以迭代为单位形成-batch","a. 以迭代为单位形成 batch",{"id":170,"text":171,"level":165},"b-跨请求批量化计算","b. 跨请求批量化计算：",{"id":173,"text":174,"level":165},"c-orca-策略-连续批次","c. Orca 策略：连续批次",{"id":176,"text":177,"level":165},"2-pd-分离优化","（2） PD 分离优化",{"id":179,"text":180,"level":165},"3-chunk-prefill-与-decode-only-batch","（3） Chunk Prefill 与 Decode-Only Batch",{"id":182,"text":182,"level":9},"模型层",{"id":184,"text":184,"level":110},"服务系统架构设计示例",{"id":186,"text":187,"level":9},"学习贴士","💡 学习贴士",{"id":189,"text":190,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]