[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-42":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_42_fm.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4corv3YmFOUMcu2G%2FQGl7IikTKQ%3D","InfiniTensor,冬季训练营,DLCompiler,高性能算子,Triton","围绕 DLCompiler 高性能算子开发实践，讲解 Triton 与九齿体系下的实现路径与优化方法。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 Triton & 九齿第六课《DLCompiler 的高性能算子开发实践》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5Hu2cwfzW8yDovvm2aLeredYQgQ%3D)\r\n\r\n\r\n\r\n2026 年 1 月 28 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 Triton&九齿系列第六课《DLCompiler 的高性能算子开发实践》开讲。\r\n\r\n本节课讲解包括大 kernel 写法、GroupMatmul 优化、千问3 Next Attention Prolog 优化以及DeepSeek V3 Attention Prolog 优化。\r\n\r\n## 基于绑核优化实现 MegaKernel 总体思路\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=82VDNO6fe%2Ftt40T9mX0aMdPRRdM%3D)\r\n\r\n## GroupMatmul的高性能实现\r\n\r\n### 1. GroupMatmul的计算流程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qhE3smmUK8gkRJj8d76zPl1%2FDak%3D)\r\n\r\n\r\n### 2. 传统实现的问题\r\n\r\n传统实现方式存在：\r\n- 通过host侧循环处理group\r\n- 产生同步开销\r\n- 计算核空闲，资源利用率低\r\n\r\n### 3. 高性能实现\r\n\r\n#### (1) 动态确定分组边界\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5tR7JyNovkUsFtL%2Ffnp38l11CPY%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y7UfeATdApbM0tEbty5V10dzoWo%3D)\r\n\r\n\r\n#### (2) 对角线分核优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wO8TxU8rExqKDIcDjvHrSPUPhag%3D)\r\n\r\n\r\n## Qwen3 Next-AttentionProlog 的高性能实现\r\n\r\n### 计算流程\r\n- 输入 x 经过 weight 矩阵得到 QQKV，经过 split 拆分\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cAZxPwNBTuJTQKnxFlM7gfDjZkE%3D)\r\n\r\n- 数据拆分过程：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aM4tya%2Fn6bQzJCIqHlVS5q1gbSs%3D)\r\n\r\n\r\n### 融合思路\r\n#### 1.融合思路\r\n- 最大化利用矩阵计算单元和向量运算单元\r\n- 避免 tensor 变换\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mRgVeeSiev1Dos3Aqsdx92r1UAc%3D)\r\n\r\n\r\n\r\n#### 2. 融合（Norm+Rope）方式一\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1YpPw7%2BVAp2tQdwOgOBaXjZ2%2FVM%3D)\r\n\r\n#### 3. 融合（Norm+Rope）方式二\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BGg2OUkMHZ6IG8b5%2B9Cjrx67ihQ%3D)\r\n\r\n#### 4. 融合 rmsnorm+sigmoid\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=huiUZgviXg15q59xUZtgPzwA3wY%3D)\r\n\r\n#### 5. 异步编程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0tjmTRMilgnFq40Ke%2Bp16racrUY%3D)\r\n\r\n#### 6. PartialROPE\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw14.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FWfshrrq10xSHbglLzYVNLhgbjE%3D)\r\n\r\n\r\n### 3. 实现效果\r\n\r\n- 通过上述优化，千问3 Next Attention Prolog 在 8 K输入场景下，相比不融合方案，速度提升约 1.8 倍。\r\n\r\n\r\n## DeepSeek V3 attention prolog的高性能实现\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw15.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xmOzZyrKVZ1FWcdWdWHHuo4wlFE%3D)\r\n\r\n\r\n### DeepSeek V3 Attention Prolog 计算流程\r\n- x 经过 weight 得到q、k、v\r\n- 独立运行 RMSNorm\r\n- split 后，部分进入 BMM，部分进入 RPE\r\n\r\n### 优化策略\r\n\r\n#### 1. 融合方式一\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw16.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qkeoa4QrPC8YGRNdwuO9dLjjTQ0%3D)\r\n\r\n\r\n#### 2. 融合方式二\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw17.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1mtUGYlgloOEIZynp2WTtsO5u10%3D)\r\n\r\n\r\n### 实现效果\r\n\r\nDeepSeek V3 的 Attention Prolog 优化后，相比单算子实现，加速比达到170%。\r\n\r\n\r\n\r\n## 总结\r\n\r\nDLCompiler 通过创新的算子开发方法和优化策略，显著提升了国产芯片上的算子性能。DLCompiler 为开发者提供了一套高效、易用的工具链，助力国产芯片在 AI 时代发挥更大价值。\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![](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-41","study-41","InfiniTensor 冬季训练营 AI 编译器第六课《Stable Diffusion 在 InfiniTensor 推理框架中的适配与工程实践》",67,"\u002Fstudy-material\u002Fstudy-42","study-42","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-43","study-43","InfiniTensor 冬季训练营 OpenCL 编程第三课《OpenCL 算子的实现与优化》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5Hu2cwfzW8yDovvm2aLeredYQgQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 28 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 Triton&九齿系列第六课《DLCompiler 的高性能算子开发实践》开讲。\u003C\u002Fp>\n\u003Cp>本节课讲解包括大 kernel 写法、GroupMatmul 优化、千问3 Next Attention Prolog 优化以及DeepSeek V3 Attention Prolog 优化。\u003C\u002Fp>\n\u003Ch2 id=\"基于绑核优化实现-megakernel-总体思路\">\u003Ca class=\"anchor\" href=\"#基于绑核优化实现-megakernel-总体思路\" aria-label=\"跳转到标题：基于绑核优化实现 MegaKernel 总体思路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>基于绑核优化实现 MegaKernel 总体思路\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=82VDNO6fe%2Ftt40T9mX0aMdPRRdM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"groupmatmul的高性能实现\">\u003Ca class=\"anchor\" href=\"#groupmatmul的高性能实现\" aria-label=\"跳转到标题：GroupMatmul的高性能实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GroupMatmul的高性能实现\u003C\u002Fh2>\u003Ch3 id=\"1-groupmatmul的计算流程\">\u003Ca class=\"anchor\" href=\"#1-groupmatmul的计算流程\" aria-label=\"跳转到标题：1. GroupMatmul的计算流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. GroupMatmul的计算流程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qhE3smmUK8gkRJj8d76zPl1%2FDak%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"2-传统实现的问题\">\u003Ca class=\"anchor\" href=\"#2-传统实现的问题\" aria-label=\"跳转到标题：2. 传统实现的问题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 传统实现的问题\u003C\u002Fh3>\u003Cp>传统实现方式存在：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>通过host侧循环处理group\u003C\u002Fli>\n\u003Cli>产生同步开销\u003C\u002Fli>\n\u003Cli>计算核空闲，资源利用率低\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"3-高性能实现\">\u003Ca class=\"anchor\" href=\"#3-高性能实现\" aria-label=\"跳转到标题：3. 高性能实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 高性能实现\u003C\u002Fh3>\u003Ch4 id=\"1-动态确定分组边界\">\u003Ca class=\"anchor\" href=\"#1-动态确定分组边界\" aria-label=\"跳转到标题：(1) 动态确定分组边界\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(1) 动态确定分组边界\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5tR7JyNovkUsFtL%2Ffnp38l11CPY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y7UfeATdApbM0tEbty5V10dzoWo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-对角线分核优化\">\u003Ca class=\"anchor\" href=\"#2-对角线分核优化\" aria-label=\"跳转到标题：(2) 对角线分核优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>(2) 对角线分核优化\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wO8TxU8rExqKDIcDjvHrSPUPhag%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"qwen3-next-attentionprolog-的高性能实现\">\u003Ca class=\"anchor\" href=\"#qwen3-next-attentionprolog-的高性能实现\" aria-label=\"跳转到标题：Qwen3 Next-AttentionProlog 的高性能实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Qwen3 Next-AttentionProlog 的高性能实现\u003C\u002Fh2>\u003Ch3 id=\"计算流程\">\u003Ca class=\"anchor\" href=\"#计算流程\" aria-label=\"跳转到标题：计算流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算流程\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cp>输入 x 经过 weight 矩阵得到 QQKV，经过 split 拆分\u003Cbr>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cAZxPwNBTuJTQKnxFlM7gfDjZkE%3D\" alt=\"\">\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>数据拆分过程：\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aM4tya%2Fn6bQzJCIqHlVS5q1gbSs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"融合思路\">\u003Ca class=\"anchor\" href=\"#融合思路\" aria-label=\"跳转到标题：融合思路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>融合思路\u003C\u002Fh3>\u003Ch4 id=\"1-融合思路\">\u003Ca class=\"anchor\" href=\"#1-融合思路\" aria-label=\"跳转到标题：1.融合思路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.融合思路\u003C\u002Fh4>\u003Cul>\n\u003Cli>最大化利用矩阵计算单元和向量运算单元\u003C\u002Fli>\n\u003Cli>避免 tensor 变换\u003Cbr>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mRgVeeSiev1Dos3Aqsdx92r1UAc%3D\" alt=\"\">\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-融合-norm-rope-方式一\">\u003Ca class=\"anchor\" href=\"#2-融合-norm-rope-方式一\" aria-label=\"跳转到标题：2. 融合（Norm+Rope）方式一\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 融合（Norm+Rope）方式一\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1YpPw7%2BVAp2tQdwOgOBaXjZ2%2FVM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-融合-norm-rope-方式二\">\u003Ca class=\"anchor\" href=\"#3-融合-norm-rope-方式二\" aria-label=\"跳转到标题：3. 融合（Norm+Rope）方式二\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 融合（Norm+Rope）方式二\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BGg2OUkMHZ6IG8b5%2B9Cjrx67ihQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"4-融合-rmsnorm-sigmoid\">\u003Ca class=\"anchor\" href=\"#4-融合-rmsnorm-sigmoid\" aria-label=\"跳转到标题：4. 融合 rmsnorm+sigmoid\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. 融合 rmsnorm+sigmoid\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=huiUZgviXg15q59xUZtgPzwA3wY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"5-异步编程\">\u003Ca class=\"anchor\" href=\"#5-异步编程\" aria-label=\"跳转到标题：5. 异步编程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>5. 异步编程\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0tjmTRMilgnFq40Ke%2Bp16racrUY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"6-partialrope\">\u003Ca class=\"anchor\" href=\"#6-partialrope\" aria-label=\"跳转到标题：6. PartialROPE\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>6. PartialROPE\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw14.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FWfshrrq10xSHbglLzYVNLhgbjE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"3-实现效果\">\u003Ca class=\"anchor\" href=\"#3-实现效果\" aria-label=\"跳转到标题：3. 实现效果\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 实现效果\u003C\u002Fh3>\u003Cul>\n\u003Cli>通过上述优化，千问3 Next Attention Prolog 在 8 K输入场景下，相比不融合方案，速度提升约 1.8 倍。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"deepseek-v3-attention-prolog的高性能实现\">\u003Ca class=\"anchor\" href=\"#deepseek-v3-attention-prolog的高性能实现\" aria-label=\"跳转到标题：DeepSeek V3 attention prolog的高性能实现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DeepSeek V3 attention prolog的高性能实现\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw15.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xmOzZyrKVZ1FWcdWdWHHuo4wlFE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"deepseek-v3-attention-prolog-计算流程\">\u003Ca class=\"anchor\" href=\"#deepseek-v3-attention-prolog-计算流程\" aria-label=\"跳转到标题：DeepSeek V3 Attention Prolog 计算流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DeepSeek V3 Attention Prolog 计算流程\u003C\u002Fh3>\u003Cul>\n\u003Cli>x 经过 weight 得到q、k、v\u003C\u002Fli>\n\u003Cli>独立运行 RMSNorm\u003C\u002Fli>\n\u003Cli>split 后，部分进入 BMM，部分进入 RPE\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"优化策略\">\u003Ca class=\"anchor\" href=\"#优化策略\" aria-label=\"跳转到标题：优化策略\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化策略\u003C\u002Fh3>\u003Ch4 id=\"1-融合方式一\">\u003Ca class=\"anchor\" href=\"#1-融合方式一\" aria-label=\"跳转到标题：1. 融合方式一\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 融合方式一\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw16.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qkeoa4QrPC8YGRNdwuO9dLjjTQ0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-融合方式二\">\u003Ca class=\"anchor\" href=\"#2-融合方式二\" aria-label=\"跳转到标题：2. 融合方式二\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 融合方式二\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_42_zw17.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1mtUGYlgloOEIZynp2WTtsO5u10%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"实现效果\">\u003Ca class=\"anchor\" href=\"#实现效果\" aria-label=\"跳转到标题：实现效果\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>实现效果\u003C\u002Fh3>\u003Cp>DeepSeek V3 的 Attention Prolog 优化后，相比单算子实现，加速比达到170%。\u003C\u002Fp>\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>DLCompiler 通过创新的算子开发方法和优化策略，显著提升了国产芯片上的算子性能。DLCompiler 为开发者提供了一套高效、易用的工具链，助力国产芯片在 AI 时代发挥更大价值。\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",[128,131,134,137,140,143,147,150,153,155,157,160,163,166,169,172,175,178,181,184,186,189,192,194,196,199],{"id":129,"text":130,"level":110},"基于绑核优化实现-megakernel-总体思路","基于绑核优化实现 MegaKernel 总体思路",{"id":132,"text":133,"level":110},"groupmatmul的高性能实现","GroupMatmul的高性能实现",{"id":135,"text":136,"level":9},"1-groupmatmul的计算流程","1. GroupMatmul的计算流程",{"id":138,"text":139,"level":9},"2-传统实现的问题","2. 传统实现的问题",{"id":141,"text":142,"level":9},"3-高性能实现","3. 高性能实现",{"id":144,"text":145,"level":146},"1-动态确定分组边界","(1) 动态确定分组边界",4,{"id":148,"text":149,"level":146},"2-对角线分核优化","(2) 对角线分核优化",{"id":151,"text":152,"level":110},"qwen3-next-attentionprolog-的高性能实现","Qwen3 Next-AttentionProlog 的高性能实现",{"id":154,"text":154,"level":9},"计算流程",{"id":156,"text":156,"level":9},"融合思路",{"id":158,"text":159,"level":146},"1-融合思路","1.融合思路",{"id":161,"text":162,"level":146},"2-融合-norm-rope-方式一","2. 融合（Norm+Rope）方式一",{"id":164,"text":165,"level":146},"3-融合-norm-rope-方式二","3. 融合（Norm+Rope）方式二",{"id":167,"text":168,"level":146},"4-融合-rmsnorm-sigmoid","4. 融合 rmsnorm+sigmoid",{"id":170,"text":171,"level":146},"5-异步编程","5. 异步编程",{"id":173,"text":174,"level":146},"6-partialrope","6. PartialROPE",{"id":176,"text":177,"level":9},"3-实现效果","3. 实现效果",{"id":179,"text":180,"level":110},"deepseek-v3-attention-prolog的高性能实现","DeepSeek V3 attention prolog的高性能实现",{"id":182,"text":183,"level":9},"deepseek-v3-attention-prolog-计算流程","DeepSeek V3 Attention Prolog 计算流程",{"id":185,"text":185,"level":9},"优化策略",{"id":187,"text":188,"level":146},"1-融合方式一","1. 融合方式一",{"id":190,"text":191,"level":146},"2-融合方式二","2. 融合方式二",{"id":193,"text":193,"level":9},"实现效果",{"id":195,"text":195,"level":110},"总结",{"id":197,"text":198,"level":9},"学习贴士","💡 学习贴士",{"id":200,"text":201,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]