[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-75":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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BwVzWBmdtkq8YPZLqXna4E1FT3M%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E67XHLL7db9GOlGHdVrBAdjqtXU%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=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ZDQtqBSR0N65mJrj3FTJxyAPi6M%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_75_fm.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lgQllrKAL%2B91LRsyTRbhuB3aF2g%3D","AI 编译器,图优化,算子融合,CUDA Graph,子图替换","讲解 AI 编译器四类图等价优化：常量折叠、冗余消除、算子融合、布局转换，介绍子图 Plugin 替换与 CUDA Graph 批捕获加速方案，覆盖推理端工程优化落地。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 AI 编译器 2：《AI 编译器中的前端优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PHzZcrlojCDiu7lu7l0pp%2BaZl6U%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 23 日，2026 夏季 InfiniTensor 训练营 AI 编译器方向第二课《AI 编译器中的前端优化》开讲。\r\n上一课《AI 编译器概述》介绍 AI 发展脉络、AI 四层技术栈，对比传统编译器与 AI 编译器差异，讲解 TVM、MLIR 两大主流框架，以及张量、Stride、计算图、张量推导等基础抽象，为本节图优化打下理论基础。\r\n\u003C\u002Fspan>\r\n\r\n本节承接上一课 AI 编译器基础概念，完整讲解四类核心等价图变换、子图替换工程实践与 CUDA Graph 底层加速方案，搭建前端图优化完整落地认知。\r\n\r\n\r\n## 图层优化基础概念\r\n### 图替换规则约束\r\n图层优化全部采用**语义等价变换**，不改变模型输入输出结果。\r\n- 模板：预先定义标准化子图匹配规则；\r\n- 图替换：匹配目标子图后内部重写，替换后的子图需要保持与原子图等价的输入输出语义。\r\n\r\n### 图层优化解决两类冗余\r\n1. **结构冗余**：无效节点、重复子图、可提前计算常量；优化手段：常量折叠、冗余节点消除、公共表达式消除；\r\n2. **读写冗余**：算子间频繁显存读写、内存排布不连续；优化手段：算子融合、数据布局转换。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G6qClhNcdhNdQQlwojHwBSx5poI%3D)\r\n\r\n## 四类核心图层优化手段\r\n### 常量折叠与常量传播\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aGrSGtvRe6jkb5Bccfp3RNYWb%2FU%3D)\r\n#### 核心作用\r\n编译期计算固定常量表达式，减少运行时 Kernel 启动与计算开销。\r\n1. 常量折叠：输入全部为常量的算子，编译期直接算出结果替换为新常量；\r\n2. 常量传播：将算出的常量向后传递，持续简化后续表达式；\r\n3. 特殊可折叠算子：Shape、Size 等仅依赖张量静态形状的运算，即便张量数值未知也可提前计算。\r\n> 实操要点：优化器会多轮反复执行该 Pass，直到计算图结构不再变化。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x8xMfBy1iUyMsc1Wx5i%2B6U1DJPY%3D)\r\n\r\n### 冗余节点消除 + 公共表达式消除\r\n#### 1.冗余节点消除（对标传统 DCE 死代码消除）\r\n删除无实际作用节点，分为四类场景：\r\n1. 节点无实际输入输出作用：单输入 Concat 等；\r\n2. 参数无意义：Cast FP32 → FP32、相同 Shape Reshape、无变化 Transpose；\r\n3. 中间过渡冗余：连续多步类型转换可合并（需保证数值等价）；\r\n4. 前后互为逆运算：连续两次 Transpose、来回 Cast 抵消。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Vnn8QcLiG97aKDQrIJNTnJCkCAA%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=juZ%2FmR6MLggKjHrUedXD57OTkQ4%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lJFutEHXsy4cSt2MG9n%2BDq0y4oM%3D)\r\n\r\n#### 2.公共表达式消除\r\n合并完全相同的重复子图，减少重复计算。\r\n判定条件：算子类型、输入、属性、数据类型、Shape 全部一致，无随机副作用才可合并。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3TLajpnUK%2BsA%2F2DzsrYVi%2FcFHss%3D)\r\n\r\n### 算子融合（最常用性能优化手段）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PoZOm7UU2V5T3j3KQjOBKNVHt1M%3D)\r\n\r\n#### 1. 核心收益\r\n合并多个细粒度算子为单个复合 Kernel，消除中间张量 Load \u002F Store 显存读写开销，降低 Kernel 启动次数。\r\n#### 2. 三类融合实现形态\r\n1. 权重吸收融合：Conv + BN、MatMul + Add，数学推导合并权重与偏置（仅推理阶段可用，训练 BN 统计量动态不可合并）；\r\n2. 自定义融合 Kernel：Gelu、Layer Norm 等复合算子单独实现高性能内核；\r\n3. 属性内嵌融合：MatMul 内置转置参数、Conv 内置 Padding，消除前置 Transpose \u002F Pad 节点。\r\n> 注意：并非融合越多越好，过度融合会拉高寄存器压力，降低硬件利用率。\r\n\r\n#### 3. 典型案例：Conv + BatchNorm 融合\r\n推理阶段 BN 均值、方差、缩放参数均为常量，通过公式推导将 BN 计算逻辑直接合并进卷积权重与偏置，省去独立 BN 算子读写。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2iY2QAXTaQTJtcyubbw1RGxPuZk%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=f%2FAk2%2FMp%2BPKo9RxodsaXY7rCX5U%3D)\r\n\r\n### 数据布局转换\r\n#### 优化逻辑：\r\n调整张量内存排布，适配硬件访存特性，实现连续向量化读取，不减少计算量但大幅降低访存耗时。\r\n1. 主流布局对比 NCHW \u002F NHWC：\r\n   - NCHW：通道分散，1×1 卷积访存不连续；\r\n   - NHWC：通道为最内层连续维度，在部分 GPU\u002FNPU 卷积计算场景下更利于向量化访问和硬件计算单元利用；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A5eGVVF0rkPO0u%2BbJut4gxtQzIk%3D)\r\n\r\n2. 异构硬件专属布局：昇腾等硬件架构，拆分通道维度贴合硬件向量宽度；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw13.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=STYhrXnGscFHK5m3TaxujyPuc4M%3D)\r\n\r\n3. 权重优化：常量卷积核离线完成布局转换，无需推理时实时 Transpose；\r\n4. 插入规则：仅在布局敏感算子前后插入 Transpose，避免频繁转换增加开销。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw14.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TkhDN0vvXLBvVwECgTyU2kZYnQY%3D)\r\n\r\n## 子图替换工程实战流程\r\n### 优化思路\r\n使用硬件 Profiler 采集算子耗时，定位占比极高、访存不连续、运行在低速单元（AI_CPU）的算子作为优化目标。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw15.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OKOfQQQB4nK8ymPtqltb7x9rsTs%3D)\r\n\r\n通用算子为兼容所有输入性能较差；若业务输入存在固定规则，可抽取整段子图，替换为定制 Plugin OP，迁移至高速计算单元执行。\r\n\r\n### 模型优化案例\r\n1. 导入原始模型，定位待优化子图输入、输出边界；\r\n2. 断开原图张量连接，插入自定义 Plugin OP；\r\n3. 删除原冗余子图节点，执行拓扑重排；\r\n4. 导出优化后模型，配套后端专用 Kernel 实现；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw16.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=duJQ4ZOzzdp%2FmTDVu%2BzNOiq%2FPYk%3D)\r\n\r\n## CUDA Graph 加速机制\r\n### 原生 Kernel Launch 痛点\r\n逐算子提交 Kernel 存在固定 CPU 调度开销（3 ~ 10 μs \u002F 次），小算子模型累计开销巨大；CPU 提交速度不足会造成 GPU 空泡、利用率低于 70%。\r\n\r\n### CUDA Graph 核心原理\r\n一次性捕获一整段计算任务拓扑，后续重复推理仅重放图，摊薄重复 Kernel 调度开销。\r\n\r\n### 完整三阶段使用流程\r\n1. 定义阶段：手动构建图 \u002F Stream 流捕获自动记录算子依赖；\r\n2. 实例化阶段：校验依赖、预分配资源，转为可执行 Graph；\r\n3. 执行阶段：调用 Graph Launch 直接重放，无需 CPU 逐个提交算子。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw17.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=joh7N4ptrwXPm0QC8ZZ6Iw%2B%2F7D0%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw18.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FIcmjtV0NTIW8IFqtVMdZoZdAEo%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw19.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=C83%2BZjKOMtojPC53VRbyEHocOUI%3D)\r\n\r\n### 适用场景与局限\r\n- 优势：静态 Decode 阶段、固定输入长度小 Batch 推理提速明显；\r\n- 局限：依赖固定张量 Shape、稳定内存地址，Prefill 动态变长场景工程复杂度更高。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw20.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hso%2FpEdDDM5RWwfC%2BN1%2By8%2Bes%2Bs%3D)\r\n\r\n\r\n\r\n\r\n## 加入训练营\r\n#### 📺完整课程内容，请观看直播或查看课程回放：\r\n**InfiniTensor 官网**： https:\u002F\u002Fwww.infinitensor.com\r\n\r\n**B站直播**： InfiniTensor 官方直播间\r\n\r\n**视频号直播**： InfiniTensor 视频号预约\r\n\r\n**答疑交流**： 训练营官方社群⬇️\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uKKZHEMMNB9PmoN2W7t8t3RWewk%3D)\r\n\r\n## 关注与交流\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2FSVGJ8eOraf6k1nlkjH0x9NZjY%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-74","study-74","2026 夏季 InfiniTensor 训练营训练系统 2：《分布式训练》",34,"\u002Fstudy-material\u002Fstudy-75","study-75","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-76","study-76","2026 夏季 InfiniTensor 训练营 CUDA 编程 3：《内存模型与规约优化》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PHzZcrlojCDiu7lu7l0pp%2BaZl6U%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 23 日，2026 夏季 InfiniTensor 训练营 AI 编译器方向第二课《AI 编译器中的前端优化》开讲。\u003Cbr>上一课《AI 编译器概述》介绍 AI 发展脉络、AI 四层技术栈，对比传统编译器与 AI 编译器差异，讲解 TVM、MLIR 两大主流框架，以及张量、Stride、计算图、张量推导等基础抽象，为本节图优化打下理论基础。\u003Cbr>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节承接上一课 AI 编译器基础概念，完整讲解四类核心等价图变换、子图替换工程实践与 CUDA Graph 底层加速方案，搭建前端图优化完整落地认知。\u003C\u002Fp>\n\u003Ch2 id=\"图层优化基础概念\">\u003Ca class=\"anchor\" href=\"#图层优化基础概念\" aria-label=\"跳转到标题：图层优化基础概念\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>图层优化基础概念\u003C\u002Fh2>\u003Ch3 id=\"图替换规则约束\">\u003Ca class=\"anchor\" href=\"#图替换规则约束\" aria-label=\"跳转到标题：图替换规则约束\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>图替换规则约束\u003C\u002Fh3>\u003Cp>图层优化全部采用\u003Cstrong>语义等价变换\u003C\u002Fstrong>，不改变模型输入输出结果。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>模板：预先定义标准化子图匹配规则；\u003C\u002Fli>\n\u003Cli>图替换：匹配目标子图后内部重写，替换后的子图需要保持与原子图等价的输入输出语义。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"图层优化解决两类冗余\">\u003Ca class=\"anchor\" href=\"#图层优化解决两类冗余\" aria-label=\"跳转到标题：图层优化解决两类冗余\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>图层优化解决两类冗余\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>结构冗余\u003C\u002Fstrong>：无效节点、重复子图、可提前计算常量；优化手段：常量折叠、冗余节点消除、公共表达式消除；\u003C\u002Fli>\n\u003Cli>\u003Cstrong>读写冗余\u003C\u002Fstrong>：算子间频繁显存读写、内存排布不连续；优化手段：算子融合、数据布局转换。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G6qClhNcdhNdQQlwojHwBSx5poI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"四类核心图层优化手段\">\u003Ca class=\"anchor\" href=\"#四类核心图层优化手段\" aria-label=\"跳转到标题：四类核心图层优化手段\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>四类核心图层优化手段\u003C\u002Fh2>\u003Ch3 id=\"常量折叠与常量传播\">\u003Ca class=\"anchor\" href=\"#常量折叠与常量传播\" aria-label=\"跳转到标题：常量折叠与常量传播\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>常量折叠与常量传播\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aGrSGtvRe6jkb5Bccfp3RNYWb%2FU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"核心作用\">\u003Ca class=\"anchor\" href=\"#核心作用\" aria-label=\"跳转到标题：核心作用\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心作用\u003C\u002Fh4>\u003Cp>编译期计算固定常量表达式，减少运行时 Kernel 启动与计算开销。\u003C\u002Fp>\n\u003Col>\n\u003Cli>常量折叠：输入全部为常量的算子，编译期直接算出结果替换为新常量；\u003C\u002Fli>\n\u003Cli>常量传播：将算出的常量向后传递，持续简化后续表达式；\u003C\u002Fli>\n\u003Cli>特殊可折叠算子：Shape、Size 等仅依赖张量静态形状的运算，即便张量数值未知也可提前计算。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cblockquote>\n\u003Cp>实操要点：优化器会多轮反复执行该 Pass，直到计算图结构不再变化。\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x8xMfBy1iUyMsc1Wx5i%2B6U1DJPY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"冗余节点消除-公共表达式消除\">\u003Ca class=\"anchor\" href=\"#冗余节点消除-公共表达式消除\" aria-label=\"跳转到标题：冗余节点消除 + 公共表达式消除\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>冗余节点消除 + 公共表达式消除\u003C\u002Fh3>\u003Ch4 id=\"1-冗余节点消除-对标传统-dce-死代码消除\">\u003Ca class=\"anchor\" href=\"#1-冗余节点消除-对标传统-dce-死代码消除\" aria-label=\"跳转到标题：1.冗余节点消除（对标传统 DCE 死代码消除）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.冗余节点消除（对标传统 DCE 死代码消除）\u003C\u002Fh4>\u003Cp>删除无实际作用节点，分为四类场景：\u003C\u002Fp>\n\u003Col>\n\u003Cli>节点无实际输入输出作用：单输入 Concat 等；\u003C\u002Fli>\n\u003Cli>参数无意义：Cast FP32 → FP32、相同 Shape Reshape、无变化 Transpose；\u003C\u002Fli>\n\u003Cli>中间过渡冗余：连续多步类型转换可合并（需保证数值等价）；\u003C\u002Fli>\n\u003Cli>前后互为逆运算：连续两次 Transpose、来回 Cast 抵消。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Vnn8QcLiG97aKDQrIJNTnJCkCAA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=juZ%2FmR6MLggKjHrUedXD57OTkQ4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lJFutEHXsy4cSt2MG9n%2BDq0y4oM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-公共表达式消除\">\u003Ca class=\"anchor\" href=\"#2-公共表达式消除\" aria-label=\"跳转到标题：2.公共表达式消除\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.公共表达式消除\u003C\u002Fh4>\u003Cp>合并完全相同的重复子图，减少重复计算。\u003Cbr>判定条件：算子类型、输入、属性、数据类型、Shape 全部一致，无随机副作用才可合并。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3TLajpnUK%2BsA%2F2DzsrYVi%2FcFHss%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"算子融合-最常用性能优化手段\">\u003Ca class=\"anchor\" href=\"#算子融合-最常用性能优化手段\" aria-label=\"跳转到标题：算子融合（最常用性能优化手段）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>算子融合（最常用性能优化手段）\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PoZOm7UU2V5T3j3KQjOBKNVHt1M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-核心收益\">\u003Ca class=\"anchor\" href=\"#1-核心收益\" aria-label=\"跳转到标题：1. 核心收益\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 核心收益\u003C\u002Fh4>\u003Cp>合并多个细粒度算子为单个复合 Kernel，消除中间张量 Load \u002F Store 显存读写开销，降低 Kernel 启动次数。\u003C\u002Fp>\n\u003Ch4 id=\"2-三类融合实现形态\">\u003Ca class=\"anchor\" href=\"#2-三类融合实现形态\" aria-label=\"跳转到标题：2. 三类融合实现形态\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 三类融合实现形态\u003C\u002Fh4>\u003Col>\n\u003Cli>权重吸收融合：Conv + BN、MatMul + Add，数学推导合并权重与偏置（仅推理阶段可用，训练 BN 统计量动态不可合并）；\u003C\u002Fli>\n\u003Cli>自定义融合 Kernel：Gelu、Layer Norm 等复合算子单独实现高性能内核；\u003C\u002Fli>\n\u003Cli>属性内嵌融合：MatMul 内置转置参数、Conv 内置 Padding，消除前置 Transpose \u002F Pad 节点。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cblockquote>\n\u003Cp>注意：并非融合越多越好，过度融合会拉高寄存器压力，降低硬件利用率。\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Ch4 id=\"3-典型案例-conv-batchnorm-融合\">\u003Ca class=\"anchor\" href=\"#3-典型案例-conv-batchnorm-融合\" aria-label=\"跳转到标题：3. 典型案例：Conv + BatchNorm 融合\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 典型案例：Conv + BatchNorm 融合\u003C\u002Fh4>\u003Cp>推理阶段 BN 均值、方差、缩放参数均为常量，通过公式推导将 BN 计算逻辑直接合并进卷积权重与偏置，省去独立 BN 算子读写。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2iY2QAXTaQTJtcyubbw1RGxPuZk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=f%2FAk2%2FMp%2BPKo9RxodsaXY7rCX5U%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"数据布局转换\">\u003Ca class=\"anchor\" href=\"#数据布局转换\" aria-label=\"跳转到标题：数据布局转换\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>数据布局转换\u003C\u002Fh3>\u003Ch4 id=\"优化逻辑\">\u003Ca class=\"anchor\" href=\"#优化逻辑\" aria-label=\"跳转到标题：优化逻辑：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化逻辑：\u003C\u002Fh4>\u003Cp>调整张量内存排布，适配硬件访存特性，实现连续向量化读取，不减少计算量但大幅降低访存耗时。\u003C\u002Fp>\n\u003Col>\n\u003Cli>主流布局对比 NCHW \u002F NHWC：\u003Cul>\n\u003Cli>NCHW：通道分散，1×1 卷积访存不连续；\u003C\u002Fli>\n\u003Cli>NHWC：通道为最内层连续维度，在部分 GPU\u002FNPU 卷积计算场景下更利于向量化访问和硬件计算单元利用；\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A5eGVVF0rkPO0u%2BbJut4gxtQzIk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>异构硬件专属布局：昇腾等硬件架构，拆分通道维度贴合硬件向量宽度；\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw13.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=STYhrXnGscFHK5m3TaxujyPuc4M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>权重优化：常量卷积核离线完成布局转换，无需推理时实时 Transpose；\u003C\u002Fli>\n\u003Cli>插入规则：仅在布局敏感算子前后插入 Transpose，避免频繁转换增加开销。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw14.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TkhDN0vvXLBvVwECgTyU2kZYnQY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"子图替换工程实战流程\">\u003Ca class=\"anchor\" href=\"#子图替换工程实战流程\" aria-label=\"跳转到标题：子图替换工程实战流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>子图替换工程实战流程\u003C\u002Fh2>\u003Ch3 id=\"优化思路\">\u003Ca class=\"anchor\" href=\"#优化思路\" aria-label=\"跳转到标题：优化思路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化思路\u003C\u002Fh3>\u003Cp>使用硬件 Profiler 采集算子耗时，定位占比极高、访存不连续、运行在低速单元（AI_CPU）的算子作为优化目标。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw15.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OKOfQQQB4nK8ymPtqltb7x9rsTs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>通用算子为兼容所有输入性能较差；若业务输入存在固定规则，可抽取整段子图，替换为定制 Plugin OP，迁移至高速计算单元执行。\u003C\u002Fp>\n\u003Ch3 id=\"模型优化案例\">\u003Ca class=\"anchor\" href=\"#模型优化案例\" aria-label=\"跳转到标题：模型优化案例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型优化案例\u003C\u002Fh3>\u003Col>\n\u003Cli>导入原始模型，定位待优化子图输入、输出边界；\u003C\u002Fli>\n\u003Cli>断开原图张量连接，插入自定义 Plugin OP；\u003C\u002Fli>\n\u003Cli>删除原冗余子图节点，执行拓扑重排；\u003C\u002Fli>\n\u003Cli>导出优化后模型，配套后端专用 Kernel 实现；\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw16.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=duJQ4ZOzzdp%2FmTDVu%2BzNOiq%2FPYk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"cuda-graph-加速机制\">\u003Ca class=\"anchor\" href=\"#cuda-graph-加速机制\" aria-label=\"跳转到标题：CUDA Graph 加速机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA Graph 加速机制\u003C\u002Fh2>\u003Ch3 id=\"原生-kernel-launch-痛点\">\u003Ca class=\"anchor\" href=\"#原生-kernel-launch-痛点\" aria-label=\"跳转到标题：原生 Kernel Launch 痛点\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>原生 Kernel Launch 痛点\u003C\u002Fh3>\u003Cp>逐算子提交 Kernel 存在固定 CPU 调度开销（3 ~ 10 μs \u002F 次），小算子模型累计开销巨大；CPU 提交速度不足会造成 GPU 空泡、利用率低于 70%。\u003C\u002Fp>\n\u003Ch3 id=\"cuda-graph-核心原理\">\u003Ca class=\"anchor\" href=\"#cuda-graph-核心原理\" aria-label=\"跳转到标题：CUDA Graph 核心原理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA Graph 核心原理\u003C\u002Fh3>\u003Cp>一次性捕获一整段计算任务拓扑，后续重复推理仅重放图，摊薄重复 Kernel 调度开销。\u003C\u002Fp>\n\u003Ch3 id=\"完整三阶段使用流程\">\u003Ca class=\"anchor\" href=\"#完整三阶段使用流程\" aria-label=\"跳转到标题：完整三阶段使用流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>完整三阶段使用流程\u003C\u002Fh3>\u003Col>\n\u003Cli>定义阶段：手动构建图 \u002F Stream 流捕获自动记录算子依赖；\u003C\u002Fli>\n\u003Cli>实例化阶段：校验依赖、预分配资源，转为可执行 Graph；\u003C\u002Fli>\n\u003Cli>执行阶段：调用 Graph Launch 直接重放，无需 CPU 逐个提交算子。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw17.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=joh7N4ptrwXPm0QC8ZZ6Iw%2B%2F7D0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw18.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FIcmjtV0NTIW8IFqtVMdZoZdAEo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw19.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=C83%2BZjKOMtojPC53VRbyEHocOUI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"适用场景与局限\">\u003Ca class=\"anchor\" href=\"#适用场景与局限\" aria-label=\"跳转到标题：适用场景与局限\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>适用场景与局限\u003C\u002Fh3>\u003Cul>\n\u003Cli>优势：静态 Decode 阶段、固定输入长度小 Batch 推理提速明显；\u003C\u002Fli>\n\u003Cli>局限：依赖固定张量 Shape、稳定内存地址，Prefill 动态变长场景工程复杂度更高。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_75_zw20.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hso%2FpEdDDM5RWwfC%2BN1%2By8%2Bes%2Bs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"加入训练营\">\u003Ca class=\"anchor\" href=\"#加入训练营\" aria-label=\"跳转到标题：加入训练营\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加入训练营\u003C\u002Fh2>\u003Ch4 id=\"完整课程内容-请观看直播或查看课程回放\">\u003Ca class=\"anchor\" href=\"#完整课程内容-请观看直播或查看课程回放\" aria-label=\"跳转到标题：📺完整课程内容，请观看直播或查看课程回放：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📺完整课程内容，请观看直播或查看课程回放：\u003C\u002Fh4>\u003Cp>\u003Cstrong>InfiniTensor 官网\u003C\u002Fstrong>： \u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\">https:\u002F\u002Fwww.infinitensor.com\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Cstrong>B站直播\u003C\u002Fstrong>： InfiniTensor 官方直播间\u003C\u002Fp>\n\u003Cp>\u003Cstrong>视频号直播\u003C\u002Fstrong>： InfiniTensor 视频号预约\u003C\u002Fp>\n\u003Cp>\u003Cstrong>答疑交流\u003C\u002Fstrong>： 训练营官方社群⬇️\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uKKZHEMMNB9PmoN2W7t8t3RWewk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"关注与交流\">\u003Ca class=\"anchor\" href=\"#关注与交流\" aria-label=\"跳转到标题：关注与交流\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关注与交流\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2FSVGJ8eOraf6k1nlkjH0x9NZjY%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,132,134,136,138,141,144,147,150,153,156,159,162,164,167,169,171,173,176,179,182,184,186,188,191],{"id":129,"text":129,"level":110},"图层优化基础概念",{"id":131,"text":131,"level":9},"图替换规则约束",{"id":133,"text":133,"level":9},"图层优化解决两类冗余",{"id":135,"text":135,"level":110},"四类核心图层优化手段",{"id":137,"text":137,"level":9},"常量折叠与常量传播",{"id":139,"text":139,"level":140},"核心作用",4,{"id":142,"text":143,"level":9},"冗余节点消除-公共表达式消除","冗余节点消除 + 公共表达式消除",{"id":145,"text":146,"level":140},"1-冗余节点消除-对标传统-dce-死代码消除","1.冗余节点消除（对标传统 DCE 死代码消除）",{"id":148,"text":149,"level":140},"2-公共表达式消除","2.公共表达式消除",{"id":151,"text":152,"level":9},"算子融合-最常用性能优化手段","算子融合（最常用性能优化手段）",{"id":154,"text":155,"level":140},"1-核心收益","1. 核心收益",{"id":157,"text":158,"level":140},"2-三类融合实现形态","2. 三类融合实现形态",{"id":160,"text":161,"level":140},"3-典型案例-conv-batchnorm-融合","3. 典型案例：Conv + BatchNorm 融合",{"id":163,"text":163,"level":9},"数据布局转换",{"id":165,"text":166,"level":140},"优化逻辑","优化逻辑：",{"id":168,"text":168,"level":110},"子图替换工程实战流程",{"id":170,"text":170,"level":9},"优化思路",{"id":172,"text":172,"level":9},"模型优化案例",{"id":174,"text":175,"level":110},"cuda-graph-加速机制","CUDA Graph 加速机制",{"id":177,"text":178,"level":9},"原生-kernel-launch-痛点","原生 Kernel Launch 痛点",{"id":180,"text":181,"level":9},"cuda-graph-核心原理","CUDA Graph 核心原理",{"id":183,"text":183,"level":9},"完整三阶段使用流程",{"id":185,"text":185,"level":9},"适用场景与局限",{"id":187,"text":187,"level":110},"加入训练营",{"id":189,"text":190,"level":140},"完整课程内容-请观看直播或查看课程回放","📺完整课程内容，请观看直播或查看课程回放：",{"id":192,"text":192,"level":110},"关注与交流"]