[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-48":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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=nCcM9ECgK5PlmraUwj%2BH%2F1S%2BELk%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kRi1yuIH5NG2docsx57MXWlzm1k%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eT9ZAOmPWlOldw%2FMHXc0X0x2%2Bhg%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_48_fm.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VXp%2Bhql3L8%2BV6xc7Snzmn25FJyw%3D","InfiniTensor,冬季训练营,Triton,九齿,天通","介绍 Triton 与九齿课程内容，围绕九齿三重天通相关机制与算子实践进行讲解。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 Triton&九齿第七课《九齿三重：天通》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9xk7vnNrtTST0dE%2FTKU9JFp0cO4%3D)\r\n\r\n\r\n\r\n\r\n2026 年 1 月 30 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 Triton&九齿系列第七课《九齿三重：天通》开讲。\r\n\r\n本节课从概念验证到性能优化，深入解析九齿与 Triton 的核心价值、性能衡量方法、实战优化案例以及配套工具链。\r\n\r\n\r\n## 九齿与 Triton 的核心价值\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bI6iYf6FGtfnx2P%2BPqp8VLozJa8%3D)\r\n\r\n\r\n### 概念区分：Kernel vs 算子\r\n\r\n在讨论算子开发时，需要明确两个关键概念：\r\n\r\n- **Kernel**：在 GPU 等设备上实际运行的底层程序，只负责计算\r\n- **算子**（Operator）：在 Kernel 基础上的上层封装，负责内存管理、参数准备等\r\n\r\n九齿和 Triton 专注于 Kernel 开发，不涉及内存管理等上层逻辑，这使得它们能够专注于计算性能优化。\r\n\r\n### 为什么需要九齿和 Triton？\r\n\r\n#### 1. 厂商库的局限性\r\n\r\n对于基础算子（如矩阵乘），厂商提供的库已经\"武装到牙齿\"，很难超越。但在以下场景中，九齿和 Triton 具有明显优势：\r\n\r\n- **新算子概念验证**：当出现新的算子需求时（如早期的 Attention 机制），可以快速实现原型\r\n- **算子融合**：将多个小算子融合为一个大算子，减少内存访问开销\r\n- **边缘场景优化**：针对特定形状或参数的算子进行专门优化\r\n\r\n#### 2. 开发效率优势\r\n\r\n- **高层次抽象**：屏蔽底层硬件细节，开发者只需关注算法逻辑\r\n- **快速原型**：相比 CUDA，开发效率提升数倍\r\n- **自动优化**：编译器自动处理内存布局、并行化等优化细节\r\n\r\n九齿和 Triton 通过加速新算子的概念验证，间接促进了新模型架构的快速迭代。\r\n\r\n## 算子性能衡量方法\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l6HbIlWZxCmMSyaRbCjGj3hCyJk%3D)\r\n\r\n\r\n### Roofline 模型\r\n\r\nRoofline 模型是衡量算子性能的重要工具，它将算子分为两类：\r\n\r\n- **访存密集型**（Memory-bound）：性能受限于内存带宽。使用 GB\u002Fs（每秒处理的字节数）衡量\r\n- **计算密集型**（Compute-bound）：性能受限于计算能力。使用 TFLOPS（每秒浮点运算次数）衡量\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LohwlL0nxCxlJFB5M66AYDJCOIQ%3D)\r\n\r\n### 矩阵乘法性能计算示例\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BB86%2BTU41EE2VWgUW0ZZDxbBbkA%3D)\r\n\r\n性能计算公式：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Bs%2B45LWWhMGDcSTQS3W51lcC6hw%3D)\r\n\r\n\r\n### RMS Norm 算子优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=U38E3J6AnqPBl7mnzuO0WuItnec%3D)\r\n\r\n#### 1. PyTorch 原生实现的问题\r\n\r\nPyTorch 组合实现 RMSNorm 实现通常由多个小算子组成：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jm4abiZCnOjLm7%2BS37hcNsZIFDY%3D)\r\n\r\n这种实现方式存在以下问题：\r\n- 多次 kernel 启动\r\n- 内存访问开销\r\n- 性能瓶颈\r\n\r\n#### 2. 九齿优化实现\r\n\r\n使用九齿将整个 RMS Norm 计算融合为单个 kernel：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9RQbqgMNSpk6ttt0EfcItbGkBMU%3D)\r\n\r\n\r\n#### 3. 性能对比结果\r\n\r\n- **算子层面**：九齿实现比 PyTorch 原生实现快数倍\r\n- **模型层面**：在整个模型运行中带来 3.5% 的性能提升\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CzOrsW5Kvk%2F4FVcyUOuCKaDA0Rk%3D)\r\n\r\n虽然 RMSNorm 在整个模型中的计算占比不高，但这种优化思路可以推广到其他算子，累积效果显著。\r\n\r\n## 九齿生态与工具链\r\n\r\n### 文档资源\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IhDdez9WCupBxWGPp0YVIHRs9Mw%3D)\r\n\r\n\r\n### 示例代码库\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rN%2BgLupQn2NM7tK%2FkxT%2BKLDbf6w%3D)\r\n\r\n\r\n\r\n九齿提供了丰富的示例代码：\r\n- **基础算子**：Add、Matrix Multiplication 等\r\n- **模型推理**：完整的模型推理示例\r\n- **性能对比**：与 PyTorch、Triton 的性能对比代码\r\n\r\n\r\n## 总结\r\n\r\n九齿和 Triton 通过高层次抽象和自动优化，在开发效率和性能之间找到了最佳平衡点。随着 AI 模型的持续演进，高效算子开发将成为 AI 系统的关键竞争力。九齿和 Triton 为开发者提供了一套完整的工具链，助力 AI 创新的快速发展。\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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NuYUa9sibpvyCfaaTlNixdKQSJk%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0AAjg4ciBP4fn4BfWUtqDIltU4c%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-47","study-47","InfiniTensor 冬季训练营推理引擎第五课《大模型服务系统》",61,"\u002Fstudy-material\u002Fstudy-48","study-48","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-49","study-49","InfiniTensor 冬季训练营推理引擎第六课《vLLM-Ascend 大模型推理》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9xk7vnNrtTST0dE%2FTKU9JFp0cO4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 30 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 Triton&九齿系列第七课《九齿三重：天通》开讲。\u003C\u002Fp>\n\u003Cp>本节课从概念验证到性能优化，深入解析九齿与 Triton 的核心价值、性能衡量方法、实战优化案例以及配套工具链。\u003C\u002Fp>\n\u003Ch2 id=\"九齿与-triton-的核心价值\">\u003Ca class=\"anchor\" href=\"#九齿与-triton-的核心价值\" aria-label=\"跳转到标题：九齿与 Triton 的核心价值\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿与 Triton 的核心价值\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bI6iYf6FGtfnx2P%2BPqp8VLozJa8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"概念区分-kernel-vs-算子\">\u003Ca class=\"anchor\" href=\"#概念区分-kernel-vs-算子\" aria-label=\"跳转到标题：概念区分：Kernel vs 算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>概念区分：Kernel vs 算子\u003C\u002Fh3>\u003Cp>在讨论算子开发时，需要明确两个关键概念：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Kernel\u003C\u002Fstrong>：在 GPU 等设备上实际运行的底层程序，只负责计算\u003C\u002Fli>\n\u003Cli>\u003Cstrong>算子\u003C\u002Fstrong>（Operator）：在 Kernel 基础上的上层封装，负责内存管理、参数准备等\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>九齿和 Triton 专注于 Kernel 开发，不涉及内存管理等上层逻辑，这使得它们能够专注于计算性能优化。\u003C\u002Fp>\n\u003Ch3 id=\"为什么需要九齿和-triton\">\u003Ca class=\"anchor\" href=\"#为什么需要九齿和-triton\" aria-label=\"跳转到标题：为什么需要九齿和 Triton？\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>为什么需要九齿和 Triton？\u003C\u002Fh3>\u003Ch4 id=\"1-厂商库的局限性\">\u003Ca class=\"anchor\" href=\"#1-厂商库的局限性\" aria-label=\"跳转到标题：1. 厂商库的局限性\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 厂商库的局限性\u003C\u002Fh4>\u003Cp>对于基础算子（如矩阵乘），厂商提供的库已经"武装到牙齿"，很难超越。但在以下场景中，九齿和 Triton 具有明显优势：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>新算子概念验证\u003C\u002Fstrong>：当出现新的算子需求时（如早期的 Attention 机制），可以快速实现原型\u003C\u002Fli>\n\u003Cli>\u003Cstrong>算子融合\u003C\u002Fstrong>：将多个小算子融合为一个大算子，减少内存访问开销\u003C\u002Fli>\n\u003Cli>\u003Cstrong>边缘场景优化\u003C\u002Fstrong>：针对特定形状或参数的算子进行专门优化\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-开发效率优势\">\u003Ca class=\"anchor\" href=\"#2-开发效率优势\" aria-label=\"跳转到标题：2. 开发效率优势\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 开发效率优势\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>高层次抽象\u003C\u002Fstrong>：屏蔽底层硬件细节，开发者只需关注算法逻辑\u003C\u002Fli>\n\u003Cli>\u003Cstrong>快速原型\u003C\u002Fstrong>：相比 CUDA，开发效率提升数倍\u003C\u002Fli>\n\u003Cli>\u003Cstrong>自动优化\u003C\u002Fstrong>：编译器自动处理内存布局、并行化等优化细节\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>九齿和 Triton 通过加速新算子的概念验证，间接促进了新模型架构的快速迭代。\u003C\u002Fp>\n\u003Ch2 id=\"算子性能衡量方法\">\u003Ca class=\"anchor\" href=\"#算子性能衡量方法\" aria-label=\"跳转到标题：算子性能衡量方法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>算子性能衡量方法\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l6HbIlWZxCmMSyaRbCjGj3hCyJk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"roofline-模型\">\u003Ca class=\"anchor\" href=\"#roofline-模型\" aria-label=\"跳转到标题：Roofline 模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Roofline 模型\u003C\u002Fh3>\u003Cp>Roofline 模型是衡量算子性能的重要工具，它将算子分为两类：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>访存密集型\u003C\u002Fstrong>（Memory-bound）：性能受限于内存带宽。使用 GB\u002Fs（每秒处理的字节数）衡量\u003C\u002Fli>\n\u003Cli>\u003Cstrong>计算密集型\u003C\u002Fstrong>（Compute-bound）：性能受限于计算能力。使用 TFLOPS（每秒浮点运算次数）衡量\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LohwlL0nxCxlJFB5M66AYDJCOIQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"矩阵乘法性能计算示例\">\u003Ca class=\"anchor\" href=\"#矩阵乘法性能计算示例\" aria-label=\"跳转到标题：矩阵乘法性能计算示例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>矩阵乘法性能计算示例\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BB86%2BTU41EE2VWgUW0ZZDxbBbkA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>性能计算公式：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Bs%2B45LWWhMGDcSTQS3W51lcC6hw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"rms-norm-算子优化\">\u003Ca class=\"anchor\" href=\"#rms-norm-算子优化\" aria-label=\"跳转到标题：RMS Norm 算子优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>RMS Norm 算子优化\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=U38E3J6AnqPBl7mnzuO0WuItnec%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-pytorch-原生实现的问题\">\u003Ca class=\"anchor\" href=\"#1-pytorch-原生实现的问题\" aria-label=\"跳转到标题：1. PyTorch 原生实现的问题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. PyTorch 原生实现的问题\u003C\u002Fh4>\u003Cp>PyTorch 组合实现 RMSNorm 实现通常由多个小算子组成：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jm4abiZCnOjLm7%2BS37hcNsZIFDY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>这种实现方式存在以下问题：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>多次 kernel 启动\u003C\u002Fli>\n\u003Cli>内存访问开销\u003C\u002Fli>\n\u003Cli>性能瓶颈\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-九齿优化实现\">\u003Ca class=\"anchor\" href=\"#2-九齿优化实现\" aria-label=\"跳转到标题：2. 九齿优化实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 九齿优化实现\u003C\u002Fh4>\u003Cp>使用九齿将整个 RMS Norm 计算融合为单个 kernel：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9RQbqgMNSpk6ttt0EfcItbGkBMU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-性能对比结果\">\u003Ca class=\"anchor\" href=\"#3-性能对比结果\" aria-label=\"跳转到标题：3. 性能对比结果\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 性能对比结果\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>算子层面\u003C\u002Fstrong>：九齿实现比 PyTorch 原生实现快数倍\u003C\u002Fli>\n\u003Cli>\u003Cstrong>模型层面\u003C\u002Fstrong>：在整个模型运行中带来 3.5% 的性能提升\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CzOrsW5Kvk%2F4FVcyUOuCKaDA0Rk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>虽然 RMSNorm 在整个模型中的计算占比不高，但这种优化思路可以推广到其他算子，累积效果显著。\u003C\u002Fp>\n\u003Ch2 id=\"九齿生态与工具链\">\u003Ca class=\"anchor\" href=\"#九齿生态与工具链\" aria-label=\"跳转到标题：九齿生态与工具链\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>九齿生态与工具链\u003C\u002Fh2>\u003Ch3 id=\"文档资源\">\u003Ca class=\"anchor\" href=\"#文档资源\" aria-label=\"跳转到标题：文档资源\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>文档资源\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IhDdez9WCupBxWGPp0YVIHRs9Mw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"示例代码库\">\u003Ca class=\"anchor\" href=\"#示例代码库\" aria-label=\"跳转到标题：示例代码库\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>示例代码库\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_48_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rN%2BgLupQn2NM7tK%2FkxT%2BKLDbf6w%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>九齿提供了丰富的示例代码：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>基础算子\u003C\u002Fstrong>：Add、Matrix Multiplication 等\u003C\u002Fli>\n\u003Cli>\u003Cstrong>模型推理\u003C\u002Fstrong>：完整的模型推理示例\u003C\u002Fli>\n\u003Cli>\u003Cstrong>性能对比\u003C\u002Fstrong>：与 PyTorch、Triton 的性能对比代码\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"总结\">\u003Ca class=\"anchor\" href=\"#总结\" aria-label=\"跳转到标题：总结\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>总结\u003C\u002Fh2>\u003Cp>九齿和 Triton 通过高层次抽象和自动优化，在开发效率和性能之间找到了最佳平衡点。随着 AI 模型的持续演进，高效算子开发将成为 AI 系统的关键竞争力。九齿和 Triton 为开发者提供了一套完整的工具链，助力 AI 创新的快速发展。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NuYUa9sibpvyCfaaTlNixdKQSJk%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0AAjg4ciBP4fn4BfWUtqDIltU4c%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,134,137,141,144,146,149,151,154,157,160,163,165,167,169,171,174],{"id":129,"text":130,"level":110},"九齿与-triton-的核心价值","九齿与 Triton 的核心价值",{"id":132,"text":133,"level":9},"概念区分-kernel-vs-算子","概念区分：Kernel vs 算子",{"id":135,"text":136,"level":9},"为什么需要九齿和-triton","为什么需要九齿和 Triton？",{"id":138,"text":139,"level":140},"1-厂商库的局限性","1. 厂商库的局限性",4,{"id":142,"text":143,"level":140},"2-开发效率优势","2. 开发效率优势",{"id":145,"text":145,"level":110},"算子性能衡量方法",{"id":147,"text":148,"level":9},"roofline-模型","Roofline 模型",{"id":150,"text":150,"level":9},"矩阵乘法性能计算示例",{"id":152,"text":153,"level":9},"rms-norm-算子优化","RMS Norm 算子优化",{"id":155,"text":156,"level":140},"1-pytorch-原生实现的问题","1. PyTorch 原生实现的问题",{"id":158,"text":159,"level":140},"2-九齿优化实现","2. 九齿优化实现",{"id":161,"text":162,"level":140},"3-性能对比结果","3. 性能对比结果",{"id":164,"text":164,"level":110},"九齿生态与工具链",{"id":166,"text":166,"level":9},"文档资源",{"id":168,"text":168,"level":9},"示例代码库",{"id":170,"text":170,"level":110},"总结",{"id":172,"text":173,"level":9},"学习贴士","💡 学习贴士",{"id":175,"text":176,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]