[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-94":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_94_fm.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=dXdM0VjolsL2BNcKho1kyDl708U%3D","九齿 DSL,算子融合,Roofline,Triton,性能调优","讲解九齿 DSL 算子融合、Roofline 性能分析法，通过 RMSNorm 实战对比 PyTorch 原生实现，配套张量可视化工具，给出算子自动调优落地流程。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 Triton & 九齿 5：《九齿三重：天通》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=23Uicmxk82R1UsMRxZRAL4kUdY4%3D)\r\n\r\n\u003Cspan style=\"font-size:85%;color:#555;\">8 月 6 日，2026 夏季 InfiniTensor 训练营Triton & 九齿方向第五课《九齿三重：天通》开讲。\r\n前置课程《九齿二重：渐悟》带领大家通过“动手写代码”的方式掌握九齿，深入理解九齿作为张量级 DSL 的核心设计理念，并通过向量加法与矩阵归约两个案例，讲解“排布”与“应用”编程范式。\u003C\u002Fspan>\r\n\r\n本节课作为九齿系列的最后一课，将跳出语法层面，深入探讨 DSL 在工业界的核心价值、性能评估方法论以及高阶可视化工具的使用，帮助大家从“会写代码”到“写好高性能算子”。\r\n\r\n## Triton 和九齿的用途\r\n### 概念辨析\r\n**1. 计算内核 (Kernel)**：纯粹的计算函数，只负责输入数据指针\u002FShape\u002FStride 到输出数据的运算，不涉及资源管理。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VqO7gh44cYiRcBEufdM4h2zca4E%3D)\r\n\r\n**2. 算子 (Operator)**：包含 Kernel + 资源分配 + 数据存储管理的完整封装（如 Class 成员变量）。\r\n\r\n### DSL 的两大核心用途\r\n#### 1. 现有算子性能优化（算子融合）：\r\n- **痛点**：在 eager execution 模式下，多个连续操作通常对应多个 Kernel，中间 Tensor 需要频繁读写 GPU Global Memory，增加访存开销。\r\n- **方案**：通过 DSL 将多个子运算融合为一个 Kernel，减少中间数据的读写开销。\r\n\r\n#### 2. 新算子\u002F新模型的概念验证 (PoC)：\r\n- **痛点**：直接使用 CUDA 开发周期长，易陷入“提前优化”陷阱。\r\n- **方案**：利用 DSL 的高开发效率快速实现 Demo，验证算法有效性后再考虑极致优化（案例：FlashAttention 系列通过重新设计 Attention 计算流程，将多个计算阶段融合，并利用 Tile 化和 IO-aware 优化降低显存访问）。\r\n\r\n## 性能评估：Roofline Model 与 Benchmark\r\n### Roofline Model 详解\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AI7jsBry%2Bdx8pMkq1%2BKXX5s%2BSnE%3D)\r\n\r\n**1. 横轴 (Arithmetic Intensity)**：计算强度（FLOPS\u002FByte）。\r\n\r\n**2. 纵轴 (Attainable Performance)**：可达性能（FLOPS\u002Fs）。\r\n\r\n**3.关键指标**：π（峰值算力）、β（带宽）。\r\n\r\n### 算子瓶颈分类\r\n**1. Memory-Bound（访存密集型）**：位于 Roofline 左侧斜坡区，受限于带宽。优化方向：减少数据搬运、提升缓存命中率。\r\n\r\n**2. Compute-Bound（计算密集型）**：位于 Roofline 右侧平台区，受限于峰值算力。优化方向：指令级并行、数学变换。\r\n\r\n**注意**：同一算子在不同 Shape 下可能处于不同区域（如窄矩阵乘可能是 Memory Bound）。\r\n\r\n### Benchmark 实战\r\n1.使用 `triton.testing.do_bench` 进行计时。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GOPGJMAcEWXLZXwvsYR%2FUzMxRFo%3D)\r\n\r\n2.**TFLOPS 计算公式推导**：\r\n- 以矩阵乘为例，$2 \times M \times N \times K$ 的来源解析（每个输出元素对应 K 次乘加）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=zGEN54e%2BVg5Z0XnPZx6jBE5bsks%3D)\r\n\r\n\r\n## 实战案例：RMSNorm 算子融合与性能提升\r\n### PyTorch 原生实现的局限\r\n1.由多个原子操作拼接而成。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JXMbFRcAi%2BTLxoQe343oUxF3ATc%3D)\r\n\r\n2.多个独立 Kernel 之间需要通过 Global Memory 传递中间结果，无法在不同计算阶段之间持续利用寄存器或 Shared Memory 保存中间数据。\r\n\r\n### 九齿融合实现\r\n1.利用 Tile 排布与 Application 组合，将所有计算逻辑封装进单一 Kernel。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qnKXjUhop%2Boq42XpKxGvREhumdU%3D)\r\n\r\n2.消除中间数据的显存搬运开销。\r\n\r\n### 性能收益数据\r\n**1. Kernel 级别**：九齿耗时仅为 PyTorch 原生的 **43.85%**。\r\n\r\n**2. 端到端推理**：在 Llama 推理场景（Input=32, Output=128）下，Token\u002Fs 从 26.71 提升至 27.65，整体吞吐提升 **3.5%**。\r\n\r\n**3. 启示**：局部算子的极致优化对系统级性能有显著累积效应。\r\n\r\n## 九齿可视化工具\r\n**1. Add 算子**：\r\n- 验证简单的 Input\u002FOther\u002FOutput 一一映射关系；切换 Program ID 观察数据块移动。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HxJ928hbO5JWRllg5mB5uqMPZ0Y%3D)\r\n\r\n**2. MM 算子**：\r\n- 利用 `target_subscript` 观察 for-loop 迭代过程中 A\u002FB 矩阵块的变化。\r\n- 验证累加器 C 的生成过程（行块 × 列块 → 部分积累加）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yfHdmDCBlp%2BaXQ%2BPD1bu%2FsiNXgs%3D)\r\n\r\n**3. Conv2d 算子**：\r\n- 处理高维数据降维可视化（Source Subscript 切片）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7H7utl7MYR93jMg3Ad5aor4rTqA%3D)\r\n\r\n**4. SDPA (Flash Attention)**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PkgjqATeW70qtWVwEbNhR%2BV2tNk%3D)\r\n\r\n\r\n\r\n## 九齿相关资源\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XBV6bIKMCUUHahpwPD5wTDViX5g%3D)\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\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\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BXH4UNzWFOrSPQr6JCB10ba6f1k%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-93","study-93","2026 夏季 InfiniTensor 训练营大咖课 1：天数智芯 AI 与加速计算负责人单天逸——《AI and AI Chip》",16,"\u002Fstudy-material\u002Fstudy-94","study-94","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-95","study-95","2026 夏季 InfiniTensor 训练营推理引擎 5：《大模型服务系统》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw01.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=23Uicmxk82R1UsMRxZRAL4kUdY4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:85%;color:#555;\">8 月 6 日，2026 夏季 InfiniTensor 训练营Triton & 九齿方向第五课《九齿三重：天通》开讲。\u003Cbr>前置课程《九齿二重：渐悟》带领大家通过“动手写代码”的方式掌握九齿，深入理解九齿作为张量级 DSL 的核心设计理念，并通过向量加法与矩阵归约两个案例，讲解“排布”与“应用”编程范式。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课作为九齿系列的最后一课，将跳出语法层面，深入探讨 DSL 在工业界的核心价值、性能评估方法论以及高阶可视化工具的使用，帮助大家从“会写代码”到“写好高性能算子”。\u003C\u002Fp>\n\u003Ch2 id=\"triton-和九齿的用途\">\u003Ca class=\"anchor\" href=\"#triton-和九齿的用途\" aria-label=\"跳转到标题：Triton 和九齿的用途\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Triton 和九齿的用途\u003C\u002Fh2>\u003Ch3 id=\"概念辨析\">\u003Ca class=\"anchor\" href=\"#概念辨析\" aria-label=\"跳转到标题：概念辨析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>概念辨析\u003C\u002Fh3>\u003Cp>\u003Cstrong>1. 计算内核 (Kernel)\u003C\u002Fstrong>：纯粹的计算函数，只负责输入数据指针\u002FShape\u002FStride 到输出数据的运算，不涉及资源管理。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw02.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VqO7gh44cYiRcBEufdM4h2zca4E%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 算子 (Operator)\u003C\u002Fstrong>：包含 Kernel + 资源分配 + 数据存储管理的完整封装（如 Class 成员变量）。\u003C\u002Fp>\n\u003Ch3 id=\"dsl-的两大核心用途\">\u003Ca class=\"anchor\" href=\"#dsl-的两大核心用途\" aria-label=\"跳转到标题：DSL 的两大核心用途\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DSL 的两大核心用途\u003C\u002Fh3>\u003Ch4 id=\"1-现有算子性能优化-算子融合\">\u003Ca class=\"anchor\" href=\"#1-现有算子性能优化-算子融合\" aria-label=\"跳转到标题：1. 现有算子性能优化（算子融合）：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 现有算子性能优化（算子融合）：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>痛点\u003C\u002Fstrong>：在 eager execution 模式下，多个连续操作通常对应多个 Kernel，中间 Tensor 需要频繁读写 GPU Global Memory，增加访存开销。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>方案\u003C\u002Fstrong>：通过 DSL 将多个子运算融合为一个 Kernel，减少中间数据的读写开销。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-新算子-新模型的概念验证-poc\">\u003Ca class=\"anchor\" href=\"#2-新算子-新模型的概念验证-poc\" aria-label=\"跳转到标题：2. 新算子\u002F新模型的概念验证 (PoC)：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 新算子\u002F新模型的概念验证 (PoC)：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>痛点\u003C\u002Fstrong>：直接使用 CUDA 开发周期长，易陷入“提前优化”陷阱。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>方案\u003C\u002Fstrong>：利用 DSL 的高开发效率快速实现 Demo，验证算法有效性后再考虑极致优化（案例：FlashAttention 系列通过重新设计 Attention 计算流程，将多个计算阶段融合，并利用 Tile 化和 IO-aware 优化降低显存访问）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"性能评估-roofline-model-与-benchmark\">\u003Ca class=\"anchor\" href=\"#性能评估-roofline-model-与-benchmark\" aria-label=\"跳转到标题：性能评估：Roofline Model 与 Benchmark\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能评估：Roofline Model 与 Benchmark\u003C\u002Fh2>\u003Ch3 id=\"roofline-model-详解\">\u003Ca class=\"anchor\" href=\"#roofline-model-详解\" aria-label=\"跳转到标题：Roofline Model 详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Roofline Model 详解\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw03.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AI7jsBry%2Bdx8pMkq1%2BKXX5s%2BSnE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. 横轴 (Arithmetic Intensity)\u003C\u002Fstrong>：计算强度（FLOPS\u002FByte）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 纵轴 (Attainable Performance)\u003C\u002Fstrong>：可达性能（FLOPS\u002Fs）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3.关键指标\u003C\u002Fstrong>：π（峰值算力）、β（带宽）。\u003C\u002Fp>\n\u003Ch3 id=\"算子瓶颈分类\">\u003Ca class=\"anchor\" href=\"#算子瓶颈分类\" aria-label=\"跳转到标题：算子瓶颈分类\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>算子瓶颈分类\u003C\u002Fh3>\u003Cp>\u003Cstrong>1. Memory-Bound（访存密集型）\u003C\u002Fstrong>：位于 Roofline 左侧斜坡区，受限于带宽。优化方向：减少数据搬运、提升缓存命中率。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Compute-Bound（计算密集型）\u003C\u002Fstrong>：位于 Roofline 右侧平台区，受限于峰值算力。优化方向：指令级并行、数学变换。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>注意\u003C\u002Fstrong>：同一算子在不同 Shape 下可能处于不同区域（如窄矩阵乘可能是 Memory Bound）。\u003C\u002Fp>\n\u003Ch3 id=\"benchmark-实战\">\u003Ca class=\"anchor\" href=\"#benchmark-实战\" aria-label=\"跳转到标题：Benchmark 实战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Benchmark 实战\u003C\u002Fh3>\u003Cp>1.使用 \u003Ccode>triton.testing.do_bench\u003C\u002Fcode> 进行计时。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw04.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GOPGJMAcEWXLZXwvsYR%2FUzMxRFo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>TFLOPS 计算公式推导\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>以矩阵乘为例，$2     imes M  imes N  imes K$ 的来源解析（每个输出元素对应 K 次乘加）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw05.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=zGEN54e%2BVg5Z0XnPZx6jBE5bsks%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"实战案例-rmsnorm-算子融合与性能提升\">\u003Ca class=\"anchor\" href=\"#实战案例-rmsnorm-算子融合与性能提升\" aria-label=\"跳转到标题：实战案例：RMSNorm 算子融合与性能提升\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>实战案例：RMSNorm 算子融合与性能提升\u003C\u002Fh2>\u003Ch3 id=\"pytorch-原生实现的局限\">\u003Ca class=\"anchor\" href=\"#pytorch-原生实现的局限\" aria-label=\"跳转到标题：PyTorch 原生实现的局限\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PyTorch 原生实现的局限\u003C\u002Fh3>\u003Cp>1.由多个原子操作拼接而成。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw06.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JXMbFRcAi%2BTLxoQe343oUxF3ATc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.多个独立 Kernel 之间需要通过 Global Memory 传递中间结果，无法在不同计算阶段之间持续利用寄存器或 Shared Memory 保存中间数据。\u003C\u002Fp>\n\u003Ch3 id=\"九齿融合实现\">\u003Ca class=\"anchor\" href=\"#九齿融合实现\" aria-label=\"跳转到标题：九齿融合实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿融合实现\u003C\u002Fh3>\u003Cp>1.利用 Tile 排布与 Application 组合，将所有计算逻辑封装进单一 Kernel。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw07.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qnKXjUhop%2Boq42XpKxGvREhumdU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.消除中间数据的显存搬运开销。\u003C\u002Fp>\n\u003Ch3 id=\"性能收益数据\">\u003Ca class=\"anchor\" href=\"#性能收益数据\" aria-label=\"跳转到标题：性能收益数据\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能收益数据\u003C\u002Fh3>\u003Cp>\u003Cstrong>1. Kernel 级别\u003C\u002Fstrong>：九齿耗时仅为 PyTorch 原生的 \u003Cstrong>43.85%\u003C\u002Fstrong>。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 端到端推理\u003C\u002Fstrong>：在 Llama 推理场景（Input=32, Output=128）下，Token\u002Fs 从 26.71 提升至 27.65，整体吞吐提升 \u003Cstrong>3.5%\u003C\u002Fstrong>。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. 启示\u003C\u002Fstrong>：局部算子的极致优化对系统级性能有显著累积效应。\u003C\u002Fp>\n\u003Ch2 id=\"九齿可视化工具\">\u003Ca class=\"anchor\" href=\"#九齿可视化工具\" aria-label=\"跳转到标题：九齿可视化工具\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿可视化工具\u003C\u002Fh2>\u003Cp>\u003Cstrong>1. Add 算子\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>验证简单的 Input\u002FOther\u002FOutput 一一映射关系；切换 Program ID 观察数据块移动。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw08.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HxJ928hbO5JWRllg5mB5uqMPZ0Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. MM 算子\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>利用 \u003Ccode>target_subscript\u003C\u002Fcode> 观察 for-loop 迭代过程中 A\u002FB 矩阵块的变化。\u003C\u002Fli>\n\u003Cli>验证累加器 C 的生成过程（行块 × 列块 → 部分积累加）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw09.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yfHdmDCBlp%2BaXQ%2BPD1bu%2FsiNXgs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Conv2d 算子\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>处理高维数据降维可视化（Source Subscript 切片）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw10.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7H7utl7MYR93jMg3Ad5aor4rTqA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. SDPA (Flash Attention)\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw11.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PkgjqATeW70qtWVwEbNhR%2BV2tNk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"九齿相关资源\">\u003Ca class=\"anchor\" href=\"#九齿相关资源\" aria-label=\"跳转到标题：九齿相关资源\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿相关资源\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_94_zw12.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XBV6bIKMCUUHahpwPD5wTDViX5g%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>\u003Cp>📺\u003Cstrong>完整课程内容，请观看直播或查看课程回放：\u003C\u002Fstrong>\u003C\u002Fp>\n\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\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792752562&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BXH4UNzWFOrSPQr6JCB10ba6f1k%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",[128,131,133,136,140,143,146,149,151,154,157,160,162,164,166,168,170],{"id":129,"text":130,"level":110},"triton-和九齿的用途","Triton 和九齿的用途",{"id":132,"text":132,"level":9},"概念辨析",{"id":134,"text":135,"level":9},"dsl-的两大核心用途","DSL 的两大核心用途",{"id":137,"text":138,"level":139},"1-现有算子性能优化-算子融合","1. 现有算子性能优化（算子融合）：",4,{"id":141,"text":142,"level":139},"2-新算子-新模型的概念验证-poc","2. 新算子\u002F新模型的概念验证 (PoC)：",{"id":144,"text":145,"level":110},"性能评估-roofline-model-与-benchmark","性能评估：Roofline Model 与 Benchmark",{"id":147,"text":148,"level":9},"roofline-model-详解","Roofline Model 详解",{"id":150,"text":150,"level":9},"算子瓶颈分类",{"id":152,"text":153,"level":9},"benchmark-实战","Benchmark 实战",{"id":155,"text":156,"level":110},"实战案例-rmsnorm-算子融合与性能提升","实战案例：RMSNorm 算子融合与性能提升",{"id":158,"text":159,"level":9},"pytorch-原生实现的局限","PyTorch 原生实现的局限",{"id":161,"text":161,"level":9},"九齿融合实现",{"id":163,"text":163,"level":9},"性能收益数据",{"id":165,"text":165,"level":110},"九齿可视化工具",{"id":167,"text":167,"level":110},"九齿相关资源",{"id":169,"text":169,"level":110},"加入训练营",{"id":171,"text":171,"level":110},"关注与交流"]