[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-109":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=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hdwJslB9xKvXwblUCm6K%2F7WK08M%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=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vmbFibCaeUQuhUDAqtxiz1%2BuTFA%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=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ljT2mPGQ2zRQVWPDQ3GbYn4x6Rw%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":120},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":44,"canonicalPath":117,"id":118,"contentType":119,"prevArticle":7},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_fm.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yWvZ44bxq6izVCkxOBRwrkBxfsE%3D","AI Agent,算子开发,MoE Kernel,上下文工程,性能优化,多范式算子","多范式算子开发实战营课程回顾，吴航讲解 AI Coding 演进、Agent 系统架构，结合 MoE Expert FFN Kernel 案例，演示 AI‑Agent 算子开发、验证与调优全流程。","2026-10-22 20:38:11",2,"课程回顾 | 多范式算子开发实战营：《基于 AI Agent 实现算子的编写、验证与优化》","9 月 22 日晚，多范式算子开发实战营邀请**联培博士生吴航老师**，带来《基于 AI Agent 实现算子的编写、验证与优化》。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw01.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lIG8d1Dvj0OdsItwUht5v3axVNQ%3D)\r\n\r\n\r\n## 一、AI Coding 的范式演进与任务边界\r\n### AI 编程的三个发展阶段\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw02.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FmeERO5R1jxaFdPlKUZoOmZGAbw%3D)\r\n\r\n**1. Predict（预测）**：基于上下文预测下一行\u002F片段，工作对象小（Token\u002F表达式级）。\r\n\r\n**2. Modify（修改）**：基于自然语言指令修改单\u002F多文件，输出为 Patch。\r\n\r\n**3. Complete（完成）**：输入 Issue\u002F仓库\u002F验收条件，输出 Commit\u002FPR\u002F分支，承担验证责任。\r\n\r\n**4.核心观点**：变化不是补全消失，而是任务范围扩大后，系统需承担更多验证责任。\r\n\r\n### 五种 Vibe Coding 工作形态\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw03.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VNxM0Bi0pNQES%2BEwilDhydGNCYc%3D)\r\n\r\n**1. Inline Completion**：光标级补全，人掌握节奏。\r\n\r\n**2. Chat\u002FAsk**：技术顾问模式，解释与分析，不改真实环境。\r\n\r\n**3. Multi-file Edit**：明确指令下的多文件 Patch 生成。\r\n\r\n**4. Local Agent**：本地仓库扫描、检索、执行命令、测试反馈。\r\n\r\n**5. Delegated Agent**：远程隔离环境后台运行，通过 PR 交付，需明确目标与验收条件。\r\n\r\n### 任务范围 vs 自治程度坐标系\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw04.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EOs7cRPH6Rn4w50Ehrs2E5idW%2FI%3D)\r\n\r\n- **左下角**：代码补全（任务小、自治低）。\r\n- **右上角**：仓库级\u002F后台 Agent（任务大、自治高）。\r\n- ⚠️右上角不代表更先进或更可靠，高自治意味着错误影响范围更大。成熟做法是**明确目标、可靠反馈、可控权限**。\r\n\r\n### 工程控制点\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw05.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lbp1jspcnyE2Twdc3%2FYe3%2Fw%2BIHo%3D)\r\n\r\n**1. Direct**：低风险、明确任务（如改阈值）。\r\n\r\n**2. Plan-First**：中等复杂度，先让 Agent 分析瓶颈假设，人确认后再改。\r\n\r\n**3. Spec-driven**：涉及调度\u002F资源边界，固定 SLO、显存预算、Fallback 行为。\r\n\r\n**4. Test-driven**：可明确测试的任务，先定义仿真\u002F回归测试场景。\r\n\r\n**5. Review-driven**：高风险性能优化，人或独立审查 Agent 策略。\r\n\r\n**6. Multi-Agent**：分工带来价值，但协调带来成本。适用于边界清晰、可独立验证的任务。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw06.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kAPSJ%2FjHtdoFguOvcoJHqmbP8lY%3D)\r\n\r\n\r\n## 二、Agent 系统架构与核心组件\r\n### 模型 ≠ Agent\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw07.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YqBfgy3f8IyI4V0pXNThColjTHc%3D)\r\n\r\n#### 1. Code Model 的最低必要原理：\r\n- 预测下一批 Token\r\n- Fill-in-the-Middle\r\n- 结构化工具调用\r\n\r\n#### 2. 三角色分工：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw08.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=X0nCGBBgMaAc4jwAdZ4QNaBQ%2FNU%3D)\r\n\r\n- **Model**：阅读信息、判断现状、选择下一步、生成参数。\r\n- **Harness**：装配上下文、校验参数、发起执行、记录状态、控制权限。\r\n- **Environment**：文件系统、Shell、Git、Benchmark、GPU 等真实反馈源。\r\n\r\n### Coding Agent Loop\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw09.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J8QPNkv5PZ9XplvVSqiIqJdWTCQ%3D)\r\n\r\n- **循环逻辑**：观察状态 → 执行动作 → 获取反馈 → 修正决策。\r\n\r\n#### 1. Context Assembly\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw10.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cw88voLXxg3%2ByjEA%2F2zXRuupupc%3D)\r\n\r\n#### 2. Tool Calling\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw11.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bordzBOPUomLNZ5f%2FSd29AXhkVE%3D)\r\n\r\n\r\n### 六层系统架构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw12.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=liCY%2BROuOyu838vYleYrSFSSlCY%3D)\r\n\r\n**1. Model & Context**：谁做决定？\r\n\r\n**2. Harness & Tools**：决定如何变成动作？\r\n\r\n**3. Runtime & Verification**：动作如何被限制和验证？\r\n\r\n**4. 应用**：当 AI 表现不佳时，逐层排查问题所在。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw13.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1SKMZWN4jq%2B1ZW43z8979wF9ULA%3D)\r\n\r\n\r\n## 三、Context Engineering 与 Harness Engineering\r\n### Context Engineering（上下文工程）\r\n\r\n**1. Prompt ≠ Context**：Prompt 关注“怎么说”，Context 关注“让模型看见什么”。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw14.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bi%2BOHv974X9LwOlihei4ey0GRes%3D)\r\n\r\n**2. 五类上下文**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw15.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aF490PACuyx2zXwV24o7Cy0ouJo%3D)\r\n\r\n**3. 生命周期**：Discover → Select → Load → Use → Compress → Persist。\r\n\r\n⚠️ Select 比 Discover 更重要；长任务需 Compaction（压缩），保留关键决策与证据，移除重复低价值内容；**动态摘要 ≠ 事实本身**。\r\n\r\n### Harness Engineering\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw16.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=osJKcX%2By5NZnWMOxbkphjT2MhBo%3D)\r\n\r\n**1. 四大组成**：指导、感知、行动工具、边界。\r\n\r\n**2. 工具设计原则**：不是越通用越好，而是**明确、结构化、边界清晰**。\r\n\r\n**3. 仓库即工作环境**：给 Agent “地图”而非“百科全书”；信息需可发现、可导航、可版本化。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw17.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HTEyFMo7UiXPmx7K0Hj0D7wVq3A%3D)\r\n\r\n**4. Executable Guardrails**：将文档规则转化为**可执行的检查\u002F测试**，产生明确失败信号，比纯文本原则更可靠。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw18.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Uhvis1MZuQHKbh9mfpdwPFOs96o%3D)\r\n\r\n### Agent Skill\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw19.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sjYvFbZP0MgxP6U9x5dylnFEYjw%3D)\r\n\r\n**1. 定义**：打包的操作方法、参考资料、验证脚本，是可版本化、可测试的团队资产。\r\n\r\n**2. 三要素**：什么时候用？怎么做？怎么证明做对了？\r\n\r\n**3. 渐进式披露**：名称\u002F描述 → `Skill.md` → 具体执行文件，避免一次性占满上下文。\r\n\r\n**4. Engineering**：像代码依赖一样审查第三方 Skill（来源、脚本行为、权限风险）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw20.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qh8Nze%2FwhaTmq5oVk1OY%2BXMKm9A%3D)\r\n\r\n## 四、评测与交付验证\r\n### Agent 评测三层次\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw21.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BxhJ6ulEZBNMbjxYQ3FXvI2MpdM%3D)\r\n\r\n**1. Outcome**：最终环境是否正确（P99 降低？吞吐提升？）。\r\n\r\n**2. Trajectory**：过程是否合理（是否建立 Baseline？读取正确模块？）。\r\n\r\n**3. System**：多次运行是否稳定（耗时、人工介入度、可复现性）。\r\n\r\n### Sandbox vs Approval\r\n- **Sandbox**：技术访问边界（文件、网络、GPU）。\r\n- **Approval**：人工授权点（高风险动作确认）。\r\n- **人的职责**：定义目标、设定边界、承担风险、最终交付判断。\r\n\r\n### 核心总结\r\n- 代码生成 ≠ 软件工程\r\n- 模型 ≠ Agent\r\n- 长上下文 ≠ Relevant Context\r\n- Skill ≠ Prompt 模板\r\n- Task Pass ≠ Delivered Software\r\n\r\n\r\n## 五、实战案例——MoE Expert FFN Kernel 优化\r\n#### 阶段 1. 重构数据通路\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw22.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=I8lSIxcExsEKnzy31wH%2BG4qkmfI%3D)\r\n\r\n- 不修改数学公式\u002FGEMM，仅调整数据组织方式。\r\n- 生产者-消费者按同一 Tile 对齐，为后续固定 Workspace 打基础。\r\n\r\n#### 阶段 2. 优化主路径\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw23.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=B29O%2Btx87xR%2F04v%2BakrxefghZjM%3D)\r\n\r\n- 拆分通用 GEMM，显式创建 TensorCoreIntrin。\r\n- Gate\u002FUp 复用输入 Token 数据，减少搬运。\r\n\r\n#### 阶段 3. 处理 Stage 1 尾块\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw24.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jgEgggSK%2Fx3r36xuJJVSdJAG1z8%3D)\r\n\r\n- 针对 MoE 专家 Token 数不均问题。\r\n- 按 actual_rows 分流：>64 走 M128，≤32 走 M32，中间走 M64。\r\n- 避免小块强行使用大 Tile 导致的空算。\r\n\r\n#### 阶段 4. 优化 Down GEMM\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw25.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G6x9pAo22QM%2BGAcLsWBB8iG0ehk%3D)\r\n\r\n- 判断 hidden-dominant vs deep-K 场景。\r\n- 不同形状匹配不同 Warp 集合，而非万能 Kernel。\r\n\r\n#### 阶段 5. 运行时工程开销优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw26.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iaPAA3IrS6RMZuTrcrWR64ndA4Y%3D)\r\n\r\n- **Kernel Cache**：相同 Shape\u002FConfig 复用 JIT 编译结果。\r\n- **Workspace Cache**：相同规格中间 Buffer 复用，避免重复分配。\r\n⚠️ Cache Key 必须完整，防止错误复用。\r\n\r\n#### 阶段 6. 优化成果\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw27.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qy0PI3PRrpSwO88uunIDg3lfplQ%3D)\r\n\r\n\r\n>至此“多范式算子开发实战营”课程内容已完结，接下来进入学员项目完成阶段，检验学习成果。\r\n\r\n\r\n\r\n## 还想进一步参与？\u003Cbr>InfiniTensor 训练营等你加入！\r\n\r\n如果你对 AI 相关知识感兴趣，欢迎关注并加入九源智能计算系统生态联合体、启元实验室主办的 **InfiniTensor 大模型与人工智能系统训练营**——一个完全免费的人工智能一站式教育训练营，每年夏季、冬季开营，从算力硬件到大模型全流程课程学习，零基础也能跟上！\r\n\r\n### 训练营亮点：\r\n• 完全免费：全程免费学习，从入门到进阶一路护航；\r\n• 门槛宽松：不限背景、不限学历、不限经验，对人工智能与大模型技术有兴趣即可报名；\r\n• 课程硬核：覆盖系统编程语言、大模型推理、算子开发、AI 编译器、性能优化等课程，还有大咖课与项目实战；\r\n• 学练结合：配套习题、算力资源与实战项目，优秀学员可获得结业证书。\r\n\r\n### 关注与参与方式：\r\n• 关注 **InfiniTensor 公众号**，第一时间获取训练营报名、课表与课程动态\r\n• 前往 **InfiniTensor 开源社区官网**（https:\u002F\u002Fwww.infinitensor.com\u002F） 注册账号、完善个人信息即可完成报名，全程免费！\r\n\r\n\r\n## 关注与交流\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j9Vsow07F8OgirV3VkJ4Rty6k7Q%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-108","study-108","课程回顾 | 多范式算子开发实战营：《基于九齿复杂算子实现案例》","\u002Fstudy-material\u002Fstudy-109","study-109","markdown",{"htmlContent":121,"headings":122},"\u003Cp>9 月 22 日晚，多范式算子开发实战营邀请\u003Cstrong>联培博士生吴航老师\u003C\u002Fstrong>，带来《基于 AI Agent 实现算子的编写、验证与优化》。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw01.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lIG8d1Dvj0OdsItwUht5v3axVNQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"一-ai-coding-的范式演进与任务边界\">\u003Ca class=\"anchor\" href=\"#一-ai-coding-的范式演进与任务边界\" aria-label=\"跳转到标题：一、AI Coding 的范式演进与任务边界\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>一、AI Coding 的范式演进与任务边界\u003C\u002Fh2>\u003Ch3 id=\"ai-编程的三个发展阶段\">\u003Ca class=\"anchor\" href=\"#ai-编程的三个发展阶段\" aria-label=\"跳转到标题：AI 编程的三个发展阶段\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>AI 编程的三个发展阶段\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw02.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FmeERO5R1jxaFdPlKUZoOmZGAbw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. Predict（预测）\u003C\u002Fstrong>：基于上下文预测下一行\u002F片段，工作对象小（Token\u002F表达式级）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Modify（修改）\u003C\u002Fstrong>：基于自然语言指令修改单\u002F多文件，输出为 Patch。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Complete（完成）\u003C\u002Fstrong>：输入 Issue\u002F仓库\u002F验收条件，输出 Commit\u002FPR\u002F分支，承担验证责任。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4.核心观点\u003C\u002Fstrong>：变化不是补全消失，而是任务范围扩大后，系统需承担更多验证责任。\u003C\u002Fp>\n\u003Ch3 id=\"五种-vibe-coding-工作形态\">\u003Ca class=\"anchor\" href=\"#五种-vibe-coding-工作形态\" aria-label=\"跳转到标题：五种 Vibe Coding 工作形态\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>五种 Vibe Coding 工作形态\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw03.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VNxM0Bi0pNQES%2BEwilDhydGNCYc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. Inline Completion\u003C\u002Fstrong>：光标级补全，人掌握节奏。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Chat\u002FAsk\u003C\u002Fstrong>：技术顾问模式，解释与分析，不改真实环境。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Multi-file Edit\u003C\u002Fstrong>：明确指令下的多文件 Patch 生成。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. Local Agent\u003C\u002Fstrong>：本地仓库扫描、检索、执行命令、测试反馈。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>5. Delegated Agent\u003C\u002Fstrong>：远程隔离环境后台运行，通过 PR 交付，需明确目标与验收条件。\u003C\u002Fp>\n\u003Ch3 id=\"任务范围-vs-自治程度坐标系\">\u003Ca class=\"anchor\" href=\"#任务范围-vs-自治程度坐标系\" aria-label=\"跳转到标题：任务范围 vs 自治程度坐标系\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>任务范围 vs 自治程度坐标系\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw04.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=EOs7cRPH6Rn4w50Ehrs2E5idW%2FI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>左下角\u003C\u002Fstrong>：代码补全（任务小、自治低）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>右上角\u003C\u002Fstrong>：仓库级\u002F后台 Agent（任务大、自治高）。\u003C\u002Fli>\n\u003Cli>⚠️右上角不代表更先进或更可靠，高自治意味着错误影响范围更大。成熟做法是\u003Cstrong>明确目标、可靠反馈、可控权限\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"工程控制点\">\u003Ca class=\"anchor\" href=\"#工程控制点\" aria-label=\"跳转到标题：工程控制点\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>工程控制点\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw05.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lbp1jspcnyE2Twdc3%2FYe3%2Fw%2BIHo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. Direct\u003C\u002Fstrong>：低风险、明确任务（如改阈值）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Plan-First\u003C\u002Fstrong>：中等复杂度，先让 Agent 分析瓶颈假设，人确认后再改。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Spec-driven\u003C\u002Fstrong>：涉及调度\u002F资源边界，固定 SLO、显存预算、Fallback 行为。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. Test-driven\u003C\u002Fstrong>：可明确测试的任务，先定义仿真\u002F回归测试场景。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>5. Review-driven\u003C\u002Fstrong>：高风险性能优化，人或独立审查 Agent 策略。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>6. Multi-Agent\u003C\u002Fstrong>：分工带来价值，但协调带来成本。适用于边界清晰、可独立验证的任务。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw06.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kAPSJ%2FjHtdoFguOvcoJHqmbP8lY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"二-agent-系统架构与核心组件\">\u003Ca class=\"anchor\" href=\"#二-agent-系统架构与核心组件\" aria-label=\"跳转到标题：二、Agent 系统架构与核心组件\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>二、Agent 系统架构与核心组件\u003C\u002Fh2>\u003Ch3 id=\"模型-agent\">\u003Ca class=\"anchor\" href=\"#模型-agent\" aria-label=\"跳转到标题：模型 ≠ Agent\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型 ≠ Agent\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw07.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YqBfgy3f8IyI4V0pXNThColjTHc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-code-model-的最低必要原理\">\u003Ca class=\"anchor\" href=\"#1-code-model-的最低必要原理\" aria-label=\"跳转到标题：1. Code Model 的最低必要原理：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Code Model 的最低必要原理：\u003C\u002Fh4>\u003Cul>\n\u003Cli>预测下一批 Token\u003C\u002Fli>\n\u003Cli>Fill-in-the-Middle\u003C\u002Fli>\n\u003Cli>结构化工具调用\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-三角色分工\">\u003Ca class=\"anchor\" href=\"#2-三角色分工\" aria-label=\"跳转到标题：2. 三角色分工：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 三角色分工：\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw08.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=X0nCGBBgMaAc4jwAdZ4QNaBQ%2FNU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Model\u003C\u002Fstrong>：阅读信息、判断现状、选择下一步、生成参数。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Harness\u003C\u002Fstrong>：装配上下文、校验参数、发起执行、记录状态、控制权限。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Environment\u003C\u002Fstrong>：文件系统、Shell、Git、Benchmark、GPU 等真实反馈源。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"coding-agent-loop\">\u003Ca class=\"anchor\" href=\"#coding-agent-loop\" aria-label=\"跳转到标题：Coding Agent Loop\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Coding Agent Loop\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw09.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J8QPNkv5PZ9XplvVSqiIqJdWTCQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>循环逻辑\u003C\u002Fstrong>：观察状态 → 执行动作 → 获取反馈 → 修正决策。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"1-context-assembly\">\u003Ca class=\"anchor\" href=\"#1-context-assembly\" aria-label=\"跳转到标题：1. Context Assembly\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Context Assembly\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw10.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cw88voLXxg3%2ByjEA%2F2zXRuupupc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-tool-calling\">\u003Ca class=\"anchor\" href=\"#2-tool-calling\" aria-label=\"跳转到标题：2. Tool Calling\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Tool Calling\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw11.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bordzBOPUomLNZ5f%2FSd29AXhkVE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"六层系统架构\">\u003Ca class=\"anchor\" href=\"#六层系统架构\" aria-label=\"跳转到标题：六层系统架构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>六层系统架构\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw12.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=liCY%2BROuOyu838vYleYrSFSSlCY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. Model & Context\u003C\u002Fstrong>：谁做决定？\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Harness & Tools\u003C\u002Fstrong>：决定如何变成动作？\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Runtime & Verification\u003C\u002Fstrong>：动作如何被限制和验证？\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. 应用\u003C\u002Fstrong>：当 AI 表现不佳时，逐层排查问题所在。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw13.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1SKMZWN4jq%2B1ZW43z8979wF9ULA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"三-context-engineering-与-harness-engineering\">\u003Ca class=\"anchor\" href=\"#三-context-engineering-与-harness-engineering\" aria-label=\"跳转到标题：三、Context Engineering 与 Harness Engineering\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>三、Context Engineering 与 Harness Engineering\u003C\u002Fh2>\u003Ch3 id=\"context-engineering-上下文工程\">\u003Ca class=\"anchor\" href=\"#context-engineering-上下文工程\" aria-label=\"跳转到标题：Context Engineering（上下文工程）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Context Engineering（上下文工程）\u003C\u002Fh3>\u003Cp>\u003Cstrong>1. Prompt ≠ Context\u003C\u002Fstrong>：Prompt 关注“怎么说”，Context 关注“让模型看见什么”。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw14.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bi%2BOHv974X9LwOlihei4ey0GRes%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 五类上下文\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw15.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aF490PACuyx2zXwV24o7Cy0ouJo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. 生命周期\u003C\u002Fstrong>：Discover → Select → Load → Use → Compress → Persist。\u003C\u002Fp>\n\u003Cp>⚠️ Select 比 Discover 更重要；长任务需 Compaction（压缩），保留关键决策与证据，移除重复低价值内容；\u003Cstrong>动态摘要 ≠ 事实本身\u003C\u002Fstrong>。\u003C\u002Fp>\n\u003Ch3 id=\"harness-engineering\">\u003Ca class=\"anchor\" href=\"#harness-engineering\" aria-label=\"跳转到标题：Harness Engineering\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Harness Engineering\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw16.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=osJKcX%2By5NZnWMOxbkphjT2MhBo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. 四大组成\u003C\u002Fstrong>：指导、感知、行动工具、边界。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 工具设计原则\u003C\u002Fstrong>：不是越通用越好，而是\u003Cstrong>明确、结构化、边界清晰\u003C\u002Fstrong>。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. 仓库即工作环境\u003C\u002Fstrong>：给 Agent “地图”而非“百科全书”；信息需可发现、可导航、可版本化。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw17.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HTEyFMo7UiXPmx7K0Hj0D7wVq3A%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. Executable Guardrails\u003C\u002Fstrong>：将文档规则转化为\u003Cstrong>可执行的检查\u002F测试\u003C\u002Fstrong>，产生明确失败信号，比纯文本原则更可靠。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw18.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Uhvis1MZuQHKbh9mfpdwPFOs96o%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"agent-skill\">\u003Ca class=\"anchor\" href=\"#agent-skill\" aria-label=\"跳转到标题：Agent Skill\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Agent Skill\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw19.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sjYvFbZP0MgxP6U9x5dylnFEYjw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. 定义\u003C\u002Fstrong>：打包的操作方法、参考资料、验证脚本，是可版本化、可测试的团队资产。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. 三要素\u003C\u002Fstrong>：什么时候用？怎么做？怎么证明做对了？\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. 渐进式披露\u003C\u002Fstrong>：名称\u002F描述 → \u003Ccode>Skill.md\u003C\u002Fcode> → 具体执行文件，避免一次性占满上下文。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>4. Engineering\u003C\u002Fstrong>：像代码依赖一样审查第三方 Skill（来源、脚本行为、权限风险）。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw20.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qh8Nze%2FwhaTmq5oVk1OY%2BXMKm9A%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"四-评测与交付验证\">\u003Ca class=\"anchor\" href=\"#四-评测与交付验证\" aria-label=\"跳转到标题：四、评测与交付验证\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>四、评测与交付验证\u003C\u002Fh2>\u003Ch3 id=\"agent-评测三层次\">\u003Ca class=\"anchor\" href=\"#agent-评测三层次\" aria-label=\"跳转到标题：Agent 评测三层次\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Agent 评测三层次\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw21.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BxhJ6ulEZBNMbjxYQ3FXvI2MpdM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cstrong>1. Outcome\u003C\u002Fstrong>：最终环境是否正确（P99 降低？吞吐提升？）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>2. Trajectory\u003C\u002Fstrong>：过程是否合理（是否建立 Baseline？读取正确模块？）。\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. System\u003C\u002Fstrong>：多次运行是否稳定（耗时、人工介入度、可复现性）。\u003C\u002Fp>\n\u003Ch3 id=\"sandbox-vs-approval\">\u003Ca class=\"anchor\" href=\"#sandbox-vs-approval\" aria-label=\"跳转到标题：Sandbox vs Approval\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Sandbox vs Approval\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>Sandbox\u003C\u002Fstrong>：技术访问边界（文件、网络、GPU）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Approval\u003C\u002Fstrong>：人工授权点（高风险动作确认）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>人的职责\u003C\u002Fstrong>：定义目标、设定边界、承担风险、最终交付判断。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"核心总结\">\u003Ca class=\"anchor\" href=\"#核心总结\" aria-label=\"跳转到标题：核心总结\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心总结\u003C\u002Fh3>\u003Cul>\n\u003Cli>代码生成 ≠ 软件工程\u003C\u002Fli>\n\u003Cli>模型 ≠ Agent\u003C\u002Fli>\n\u003Cli>长上下文 ≠ Relevant Context\u003C\u002Fli>\n\u003Cli>Skill ≠ Prompt 模板\u003C\u002Fli>\n\u003Cli>Task Pass ≠ Delivered Software\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"五-实战案例-moe-expert-ffn-kernel-优化\">\u003Ca class=\"anchor\" href=\"#五-实战案例-moe-expert-ffn-kernel-优化\" aria-label=\"跳转到标题：五、实战案例——MoE Expert FFN Kernel 优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>五、实战案例——MoE Expert FFN Kernel 优化\u003C\u002Fh2>\u003Ch4 id=\"阶段-1-重构数据通路\">\u003Ca class=\"anchor\" href=\"#阶段-1-重构数据通路\" aria-label=\"跳转到标题：阶段 1. 重构数据通路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 1. 重构数据通路\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw22.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=I8lSIxcExsEKnzy31wH%2BG4qkmfI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>不修改数学公式\u002FGEMM，仅调整数据组织方式。\u003C\u002Fli>\n\u003Cli>生产者-消费者按同一 Tile 对齐，为后续固定 Workspace 打基础。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"阶段-2-优化主路径\">\u003Ca class=\"anchor\" href=\"#阶段-2-优化主路径\" aria-label=\"跳转到标题：阶段 2. 优化主路径\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 2. 优化主路径\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw23.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=B29O%2Btx87xR%2F04v%2BakrxefghZjM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>拆分通用 GEMM，显式创建 TensorCoreIntrin。\u003C\u002Fli>\n\u003Cli>Gate\u002FUp 复用输入 Token 数据，减少搬运。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"阶段-3-处理-stage-1-尾块\">\u003Ca class=\"anchor\" href=\"#阶段-3-处理-stage-1-尾块\" aria-label=\"跳转到标题：阶段 3. 处理 Stage 1 尾块\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 3. 处理 Stage 1 尾块\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw24.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jgEgggSK%2Fx3r36xuJJVSdJAG1z8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>针对 MoE 专家 Token 数不均问题。\u003C\u002Fli>\n\u003Cli>按 actual_rows 分流：>64 走 M128，≤32 走 M32，中间走 M64。\u003C\u002Fli>\n\u003Cli>避免小块强行使用大 Tile 导致的空算。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"阶段-4-优化-down-gemm\">\u003Ca class=\"anchor\" href=\"#阶段-4-优化-down-gemm\" aria-label=\"跳转到标题：阶段 4. 优化 Down GEMM\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 4. 优化 Down GEMM\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw25.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G6x9pAo22QM%2BGAcLsWBB8iG0ehk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>判断 hidden-dominant vs deep-K 场景。\u003C\u002Fli>\n\u003Cli>不同形状匹配不同 Warp 集合，而非万能 Kernel。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"阶段-5-运行时工程开销优化\">\u003Ca class=\"anchor\" href=\"#阶段-5-运行时工程开销优化\" aria-label=\"跳转到标题：阶段 5. 运行时工程开销优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 5. 运行时工程开销优化\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw26.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iaPAA3IrS6RMZuTrcrWR64ndA4Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Kernel Cache\u003C\u002Fstrong>：相同 Shape\u002FConfig 复用 JIT 编译结果。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Workspace Cache\u003C\u002Fstrong>：相同规格中间 Buffer 复用，避免重复分配。\u003Cbr>⚠️ Cache Key 必须完整，防止错误复用。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"阶段-6-优化成果\">\u003Ca class=\"anchor\" href=\"#阶段-6-优化成果\" aria-label=\"跳转到标题：阶段 6. 优化成果\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>阶段 6. 优化成果\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_109_zw27.jpg?Expires=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Qy0PI3PRrpSwO88uunIDg3lfplQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>至此“多范式算子开发实战营”课程内容已完结，接下来进入学员项目完成阶段，检验学习成果。\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Ch2 id=\"还想进一步参与-infinitensor-训练营等你加入\">\u003Ca class=\"anchor\" href=\"#还想进一步参与-infinitensor-训练营等你加入\" aria-label=\"跳转到标题：还想进一步参与？InfiniTensor 训练营等你加入！\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>还想进一步参与？\u003Cbr>InfiniTensor 训练营等你加入！\u003C\u002Fh2>\u003Cp>如果你对 AI 相关知识感兴趣，欢迎关注并加入九源智能计算系统生态联合体、启元实验室主办的 \u003Cstrong>InfiniTensor 大模型与人工智能系统训练营\u003C\u002Fstrong>——一个完全免费的人工智能一站式教育训练营，每年夏季、冬季开营，从算力硬件到大模型全流程课程学习，零基础也能跟上！\u003C\u002Fp>\n\u003Ch3 id=\"训练营亮点\">\u003Ca class=\"anchor\" href=\"#训练营亮点\" aria-label=\"跳转到标题：训练营亮点：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>训练营亮点：\u003C\u002Fh3>\u003Cp>• 完全免费：全程免费学习，从入门到进阶一路护航；\u003Cbr>• 门槛宽松：不限背景、不限学历、不限经验，对人工智能与大模型技术有兴趣即可报名；\u003Cbr>• 课程硬核：覆盖系统编程语言、大模型推理、算子开发、AI 编译器、性能优化等课程，还有大咖课与项目实战；\u003Cbr>• 学练结合：配套习题、算力资源与实战项目，优秀学员可获得结业证书。\u003C\u002Fp>\n\u003Ch3 id=\"关注与参与方式\">\u003Ca class=\"anchor\" href=\"#关注与参与方式\" aria-label=\"跳转到标题：关注与参与方式：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关注与参与方式：\u003C\u002Fh3>\u003Cp>• 关注 \u003Cstrong>InfiniTensor 公众号\u003C\u002Fstrong>，第一时间获取训练营报名、课表与课程动态\u003Cbr>• 前往 \u003Cstrong>InfiniTensor 开源社区官网\u003C\u002Fstrong>（\u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\u002F%EF%BC%89\">https:\u002F\u002Fwww.infinitensor.com\u002F）\u003C\u002Fa> 注册账号、完善个人信息即可完成报名，全程免费！\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=1792750072&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j9Vsow07F8OgirV3VkJ4Rty6k7Q%3D\" alt=\"\">\u003C\u002Fp>\n",[123,126,129,132,135,137,140,143,147,150,153,156,159,161,164,167,170,173,176,179,182,184,187,190,193,196,199,202,205,208,211,214],{"id":124,"text":125,"level":110},"一-ai-coding-的范式演进与任务边界","一、AI Coding 的范式演进与任务边界",{"id":127,"text":128,"level":9},"ai-编程的三个发展阶段","AI 编程的三个发展阶段",{"id":130,"text":131,"level":9},"五种-vibe-coding-工作形态","五种 Vibe Coding 工作形态",{"id":133,"text":134,"level":9},"任务范围-vs-自治程度坐标系","任务范围 vs 自治程度坐标系",{"id":136,"text":136,"level":9},"工程控制点",{"id":138,"text":139,"level":110},"二-agent-系统架构与核心组件","二、Agent 系统架构与核心组件",{"id":141,"text":142,"level":9},"模型-agent","模型 ≠ Agent",{"id":144,"text":145,"level":146},"1-code-model-的最低必要原理","1. Code Model 的最低必要原理：",4,{"id":148,"text":149,"level":146},"2-三角色分工","2. 三角色分工：",{"id":151,"text":152,"level":9},"coding-agent-loop","Coding Agent Loop",{"id":154,"text":155,"level":146},"1-context-assembly","1. Context Assembly",{"id":157,"text":158,"level":146},"2-tool-calling","2. Tool Calling",{"id":160,"text":160,"level":9},"六层系统架构",{"id":162,"text":163,"level":110},"三-context-engineering-与-harness-engineering","三、Context Engineering 与 Harness Engineering",{"id":165,"text":166,"level":9},"context-engineering-上下文工程","Context Engineering（上下文工程）",{"id":168,"text":169,"level":9},"harness-engineering","Harness Engineering",{"id":171,"text":172,"level":9},"agent-skill","Agent Skill",{"id":174,"text":175,"level":110},"四-评测与交付验证","四、评测与交付验证",{"id":177,"text":178,"level":9},"agent-评测三层次","Agent 评测三层次",{"id":180,"text":181,"level":9},"sandbox-vs-approval","Sandbox vs Approval",{"id":183,"text":183,"level":9},"核心总结",{"id":185,"text":186,"level":110},"五-实战案例-moe-expert-ffn-kernel-优化","五、实战案例——MoE Expert FFN Kernel 优化",{"id":188,"text":189,"level":146},"阶段-1-重构数据通路","阶段 1. 重构数据通路",{"id":191,"text":192,"level":146},"阶段-2-优化主路径","阶段 2. 优化主路径",{"id":194,"text":195,"level":146},"阶段-3-处理-stage-1-尾块","阶段 3. 处理 Stage 1 尾块",{"id":197,"text":198,"level":146},"阶段-4-优化-down-gemm","阶段 4. 优化 Down GEMM",{"id":200,"text":201,"level":146},"阶段-5-运行时工程开销优化","阶段 5. 运行时工程开销优化",{"id":203,"text":204,"level":146},"阶段-6-优化成果","阶段 6. 优化成果",{"id":206,"text":207,"level":110},"还想进一步参与-infinitensor-训练营等你加入","还想进一步参与？InfiniTensor 训练营等你加入！",{"id":209,"text":210,"level":9},"训练营亮点","训练营亮点：",{"id":212,"text":213,"level":9},"关注与参与方式","关注与参与方式：",{"id":215,"text":215,"level":110},"关注与交流"]