[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-101":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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VJSZCyae51iopwoi0izQIKD64uI%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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wHs6e9DQam6eCjdRCa17%2BAxk2V8%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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JKqlkIO2CSuyOH0jhDR9YJP0gYU%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_101_fm.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9N%2FgByTs%2F%2F38gr%2FV99mj0qWkgNY%3D","InfiniLM,大模型推理,LLAI SYS,算子开发,异构加速","2026 夏季 InfiniTensor 训练营《InfiniLM》详解 LLAI SYS 作业工程结构，解析 InfiniLM\u002FInfiniCore 项目架构，提供 MLA、MTP、MoE 等新模型适配、性能优化、量化推理等选题指南，明确作业与项目提交规范及评审标准。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营推理引擎 6：《InfiniLM》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw01.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BPQ9LG3CbM9GxuuXNGt%2BvVDP1g0%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">8 月 11 日，2026 夏季 InfiniTensor 训练营推理引擎方向第六课《InfiniLM》开讲。\r\n前置课程《大模型服务系统》讲了对话前端、OpenAI 标准接口、KV Cache 内存管理、批量调度等内容\u003C\u002Fspan>\r\n\r\n本节课是推理方向最后一节课，聚焦 **InfiniLM**，解析作业与项目所需的工程架构，介绍实现路径与评审标准，并解析 InfiniCore 与 InfiniLM。\r\n\r\n## 关键时间节点（Deadline）\r\n\r\n请务必关注以下时间点，逾期提交可能影响评审：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw02.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rfzlWlxEXd9kCbzw%2Fs9odJD3GmE%3D)\r\n\r\n\r\n## 作业工程 LLAI SYS 解析\r\n\r\n**LLAI SYS (Let’s Learn AI System)** 是专为教学设计的轻量级推理框架，集成了数据调度、资源管理与算子实现，便于理解全貌。\r\n\r\n### 项目结构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw03.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5PDk%2FAKLHMdAk3u7d1iYBBqzHbQ%3D)\r\n\r\n#### include 核心模块\r\n- **张量 (`tensor.h`)**：生命周期管理、信息获取、视图变换、维度重排。\r\n- **运行时 (`runtime.h`)**：设备生命周期、启动流程、流、内存管理。\r\n- **算子 (`ops.h`)**：RMS_Norm, RoPE, Softmax, Linear , Add , Argmax 等。\r\n- **模型 (Models)**：Qwen2-1.5B 推理实现，包括权重加载、配置读取、KV Cache 管理及推理生命周期。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw04.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mwQhV3%2FG03lCrzCNtJwUpsKcepc%3D)\r\n\r\n\r\n## 作业具体要求\r\n### Task 0: 环境准备\r\n- 安装编译器与基础依赖库。\r\n- Fork 指定个人仓库，构建并运行基础测试。\r\n- 下载指定模型。\r\n\r\n### Task 1: 张量\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw05.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ratd2bepG2I4nh2f58HDbTWVL78%3D)\r\n\r\n\r\n### Task 2: 算子\r\n- 必做：Argmax，Embedding, Linear, RMS_Norm, RoPE, Attention, Swiglu 。\r\n- 选做：Rearrange（张量重排）。\r\n\r\n### Task 3: Qwen-1.5B 文本生成\r\n- Python 前端 (`python\u002Fqwen2.py`) + C++ 后端实现。\r\n- 完善 `ctypes` 封装与 KV Cache 管理。\r\n- 通过 `python test\u002Ftest_infer.py` 验证（CPU 推理速度较慢属正常现象）。\r\n\r\n### Task 4: 异构加速适配（重点）\r\n> **新要求**：作业阶段即要求实现硬件加速，作为项目晋级筛选门槛。\r\n\r\n- **平台要求**：NVIDIA (必选) + 天数智芯\u002F摩尔线程\u002F沐曦 (三选一)。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw06.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FgUCT0UfhyYnGKgMKv6YB4cr68M%3D)\r\n\r\n## 项目阶段选题指南\r\n\r\n**核心原则**：为 **InfiniLM** 主仓库贡献**有意义、干净、独立**的代码。避免喧宾夺主或过度侵入式修改。\r\n\r\n### 选题方向推荐\r\n#### 1. 新模型\u002F新结构适配（高价值）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw07.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LHKQ6I1x4ebeKVqnuUx3NgqqNZk%3D)\r\n\r\n- **MLA (Multi-head Latent Attention)**：DeepSeek 系列结构，通过低秩压缩显著降低 KV Cache 的存储开销。（*注：勿直接复用旧版 DeepSeek V2 代码，建议基于新接口实现。*）\r\n- **MTP (Multi-token Prediction)**：通过额外预测后续多个 Token，为投机解码等推理加速方案提供基础。参考 MiniCPM4 + MiniCPM-1Go 实现。\r\n- **MoE (Mixture of Experts)**：性能优化、新平台适配或引入新 MoE 模型（如 Qwen3-MoE, DeepSeek-V2）。\r\n- **NSA \u002F Mamba \u002F RWKV**：长文本优势结构或 SSM 类模型。\r\n- **高难度模型**：UltraMem, Titans, Minimax（基础分更高）。\r\n\r\n\r\n#### 2. 性能优化\r\n- **离线优化**：算子实现、调度策略。\r\n- **服务优化**：降低空泡、高效计算。\r\n- ⚠️ **注意**：控制改动范围，避免侵入式修改导致验证成本过高。\r\n\r\n#### 3. 框架与功能完善\r\n- 流式输出\r\n- Response API 支持\r\n- 诊断工具\r\n\r\n#### 4. 量化与低精度推理\r\n- 权重量化、KV Cache 量化\u002F压缩、混合精度执行。\r\n- 目标：良好的量化兼容性（如 W8A8 无缝切换）。\r\n\r\n#### 5. 可靠性与工程工具\r\n- Profiler 辅助工具（尤其针对缺乏 Nsight 的国产平台）。\r\n- 错误诊断与性能采集工具。\r\n\r\n### 提交与评判\r\n1. **提交目标**：\r\n   - 作业 → Fork 的个人仓库\r\n   - 项目 → InfiniLM 主仓库\r\n2. **报告规范**：PR Comment 或 Markdown 文件，必须包含**清晰复现步骤**。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw08.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kK9GglvVpJeE4xSL6vmE41EqHnI%3D)\r\n\r\n## 项目工程 InfiniLM\u002FInfiniCore 概览\r\n### InfiniCore (底层支撑)\r\n-   **InfiniOP**：提供算子。\r\n-   **InfiniCCL**：通信库（AllReduce, Send\u002FRecv for PP）。\r\n-   **InfiniRT**：运行时、CUDA Graph 支持。\r\n-   **InfiniCore 封装层**：对底层算子及运行时能力进行封装，提供更简洁的调用接口，并通过缓存机制复用相关资源。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw09.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Tp%2BfnzdObCPdfvy%2B9OHOhgKw0W4%3D)\r\n\r\n### InfiniLM (推理框架)\r\n- **Pybind11**：Python\u002FC++ 绑定。\r\n- **Engine**：基础引擎、分布式 Worker、边界保护。\r\n- **Layers**：通用模块（Attention, Linear, MLP, Quantization）。\r\n- **Models**：模块化模型实现（复用 Layers，纯添加式修改）。\r\n- **Processor (Python)**：文本处理、EOS\u002FChatTemplate 隔离封装。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw10.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2Cdy9keEpffqPDBra7rWEViwIgM%3D)\r\n\r\n\r\n> 推理方向的征程已进入实战阶段。希望大家利用好 LLAI SYS 夯实基础，在 InfiniLM 项目中展现创造力与工程能力。预祝各位在作业与项目中取得优异成绩！\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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=dUe%2BM0SLl6KCrp97rPdQhphT%2Bck%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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6FVLIDG%2FuTYOS25cEBcGpul%2Bzno%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-100","study-100","2026 夏季 InfiniTensor 训练营 AI 编程与智能体应用 9：《麒麟 Agent 运行时系统和智算软件栈优化方法》",9,"\u002Fstudy-material\u002Fstudy-101","study-101","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-102","study-102","2026 夏季 InfiniTensor 训练营前沿模型 5：《多模态智能及其应用》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw01.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BPQ9LG3CbM9GxuuXNGt%2BvVDP1g0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">8 月 11 日，2026 夏季 InfiniTensor 训练营推理引擎方向第六课《InfiniLM》开讲。\u003Cbr>前置课程《大模型服务系统》讲了对话前端、OpenAI 标准接口、KV Cache 内存管理、批量调度等内容\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课是推理方向最后一节课，聚焦 \u003Cstrong>InfiniLM\u003C\u002Fstrong>，解析作业与项目所需的工程架构，介绍实现路径与评审标准，并解析 InfiniCore 与 InfiniLM。\u003C\u002Fp>\n\u003Ch2 id=\"关键时间节点-deadline\">\u003Ca class=\"anchor\" href=\"#关键时间节点-deadline\" aria-label=\"跳转到标题：关键时间节点（Deadline）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关键时间节点（Deadline）\u003C\u002Fh2>\u003Cp>请务必关注以下时间点，逾期提交可能影响评审：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw02.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rfzlWlxEXd9kCbzw%2Fs9odJD3GmE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"作业工程-llai-sys-解析\">\u003Ca class=\"anchor\" href=\"#作业工程-llai-sys-解析\" aria-label=\"跳转到标题：作业工程 LLAI SYS 解析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>作业工程 LLAI SYS 解析\u003C\u002Fh2>\u003Cp>\u003Cstrong>LLAI SYS (Let’s Learn AI System)\u003C\u002Fstrong> 是专为教学设计的轻量级推理框架，集成了数据调度、资源管理与算子实现，便于理解全貌。\u003C\u002Fp>\n\u003Ch3 id=\"项目结构\">\u003Ca class=\"anchor\" href=\"#项目结构\" aria-label=\"跳转到标题：项目结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>项目结构\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw03.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5PDk%2FAKLHMdAk3u7d1iYBBqzHbQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"include-核心模块\">\u003Ca class=\"anchor\" href=\"#include-核心模块\" aria-label=\"跳转到标题：include 核心模块\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>include 核心模块\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>张量 (\u003Ccode>tensor.h\u003C\u002Fcode>)\u003C\u002Fstrong>：生命周期管理、信息获取、视图变换、维度重排。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>运行时 (\u003Ccode>runtime.h\u003C\u002Fcode>)\u003C\u002Fstrong>：设备生命周期、启动流程、流、内存管理。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>算子 (\u003Ccode>ops.h\u003C\u002Fcode>)\u003C\u002Fstrong>：RMS_Norm, RoPE, Softmax, Linear , Add , Argmax 等。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>模型 (Models)\u003C\u002Fstrong>：Qwen2-1.5B 推理实现，包括权重加载、配置读取、KV Cache 管理及推理生命周期。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw04.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mwQhV3%2FG03lCrzCNtJwUpsKcepc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"作业具体要求\">\u003Ca class=\"anchor\" href=\"#作业具体要求\" aria-label=\"跳转到标题：作业具体要求\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>作业具体要求\u003C\u002Fh2>\u003Ch3 id=\"task-0-环境准备\">\u003Ca class=\"anchor\" href=\"#task-0-环境准备\" aria-label=\"跳转到标题：Task 0: 环境准备\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Task 0: 环境准备\u003C\u002Fh3>\u003Cul>\n\u003Cli>安装编译器与基础依赖库。\u003C\u002Fli>\n\u003Cli>Fork 指定个人仓库，构建并运行基础测试。\u003C\u002Fli>\n\u003Cli>下载指定模型。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"task-1-张量\">\u003Ca class=\"anchor\" href=\"#task-1-张量\" aria-label=\"跳转到标题：Task 1: 张量\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Task 1: 张量\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw05.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ratd2bepG2I4nh2f58HDbTWVL78%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"task-2-算子\">\u003Ca class=\"anchor\" href=\"#task-2-算子\" aria-label=\"跳转到标题：Task 2: 算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Task 2: 算子\u003C\u002Fh3>\u003Cul>\n\u003Cli>必做：Argmax，Embedding, Linear, RMS_Norm, RoPE, Attention, Swiglu 。\u003C\u002Fli>\n\u003Cli>选做：Rearrange（张量重排）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"task-3-qwen-1-5b-文本生成\">\u003Ca class=\"anchor\" href=\"#task-3-qwen-1-5b-文本生成\" aria-label=\"跳转到标题：Task 3: Qwen-1.5B 文本生成\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Task 3: Qwen-1.5B 文本生成\u003C\u002Fh3>\u003Cul>\n\u003Cli>Python 前端 (\u003Ccode>python\u002Fqwen2.py\u003C\u002Fcode>) + C++ 后端实现。\u003C\u002Fli>\n\u003Cli>完善 \u003Ccode>ctypes\u003C\u002Fcode> 封装与 KV Cache 管理。\u003C\u002Fli>\n\u003Cli>通过 \u003Ccode>python test\u002Ftest_infer.py\u003C\u002Fcode> 验证（CPU 推理速度较慢属正常现象）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"task-4-异构加速适配-重点\">\u003Ca class=\"anchor\" href=\"#task-4-异构加速适配-重点\" aria-label=\"跳转到标题：Task 4: 异构加速适配（重点）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Task 4: 异构加速适配（重点）\u003C\u002Fh3>\u003Cblockquote>\n\u003Cp>\u003Cstrong>新要求\u003C\u002Fstrong>：作业阶段即要求实现硬件加速，作为项目晋级筛选门槛。\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Cul>\n\u003Cli>\u003Cstrong>平台要求\u003C\u002Fstrong>：NVIDIA (必选) + 天数智芯\u002F摩尔线程\u002F沐曦 (三选一)。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw06.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FgUCT0UfhyYnGKgMKv6YB4cr68M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"项目阶段选题指南\">\u003Ca class=\"anchor\" href=\"#项目阶段选题指南\" aria-label=\"跳转到标题：项目阶段选题指南\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>项目阶段选题指南\u003C\u002Fh2>\u003Cp>\u003Cstrong>核心原则\u003C\u002Fstrong>：为 \u003Cstrong>InfiniLM\u003C\u002Fstrong> 主仓库贡献\u003Cstrong>有意义、干净、独立\u003C\u002Fstrong>的代码。避免喧宾夺主或过度侵入式修改。\u003C\u002Fp>\n\u003Ch3 id=\"选题方向推荐\">\u003Ca class=\"anchor\" href=\"#选题方向推荐\" aria-label=\"跳转到标题：选题方向推荐\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>选题方向推荐\u003C\u002Fh3>\u003Ch4 id=\"1-新模型-新结构适配-高价值\">\u003Ca class=\"anchor\" href=\"#1-新模型-新结构适配-高价值\" aria-label=\"跳转到标题：1. 新模型\u002F新结构适配（高价值）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 新模型\u002F新结构适配（高价值）\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw07.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LHKQ6I1x4ebeKVqnuUx3NgqqNZk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>MLA (Multi-head Latent Attention)\u003C\u002Fstrong>：DeepSeek 系列结构，通过低秩压缩显著降低 KV Cache 的存储开销。（\u003Cem>注：勿直接复用旧版 DeepSeek V2 代码，建议基于新接口实现。\u003C\u002Fem>）\u003C\u002Fli>\n\u003Cli>\u003Cstrong>MTP (Multi-token Prediction)\u003C\u002Fstrong>：通过额外预测后续多个 Token，为投机解码等推理加速方案提供基础。参考 MiniCPM4 + MiniCPM-1Go 实现。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>MoE (Mixture of Experts)\u003C\u002Fstrong>：性能优化、新平台适配或引入新 MoE 模型（如 Qwen3-MoE, DeepSeek-V2）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>NSA \u002F Mamba \u002F RWKV\u003C\u002Fstrong>：长文本优势结构或 SSM 类模型。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>高难度模型\u003C\u002Fstrong>：UltraMem, Titans, Minimax（基础分更高）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-性能优化\">\u003Ca class=\"anchor\" href=\"#2-性能优化\" aria-label=\"跳转到标题：2. 性能优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 性能优化\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>离线优化\u003C\u002Fstrong>：算子实现、调度策略。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>服务优化\u003C\u002Fstrong>：降低空泡、高效计算。\u003C\u002Fli>\n\u003Cli>⚠️ \u003Cstrong>注意\u003C\u002Fstrong>：控制改动范围，避免侵入式修改导致验证成本过高。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-框架与功能完善\">\u003Ca class=\"anchor\" href=\"#3-框架与功能完善\" aria-label=\"跳转到标题：3. 框架与功能完善\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 框架与功能完善\u003C\u002Fh4>\u003Cul>\n\u003Cli>流式输出\u003C\u002Fli>\n\u003Cli>Response API 支持\u003C\u002Fli>\n\u003Cli>诊断工具\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"4-量化与低精度推理\">\u003Ca class=\"anchor\" href=\"#4-量化与低精度推理\" aria-label=\"跳转到标题：4. 量化与低精度推理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. 量化与低精度推理\u003C\u002Fh4>\u003Cul>\n\u003Cli>权重量化、KV Cache 量化\u002F压缩、混合精度执行。\u003C\u002Fli>\n\u003Cli>目标：良好的量化兼容性（如 W8A8 无缝切换）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"5-可靠性与工程工具\">\u003Ca class=\"anchor\" href=\"#5-可靠性与工程工具\" aria-label=\"跳转到标题：5. 可靠性与工程工具\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>5. 可靠性与工程工具\u003C\u002Fh4>\u003Cul>\n\u003Cli>Profiler 辅助工具（尤其针对缺乏 Nsight 的国产平台）。\u003C\u002Fli>\n\u003Cli>错误诊断与性能采集工具。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"提交与评判\">\u003Ca class=\"anchor\" href=\"#提交与评判\" aria-label=\"跳转到标题：提交与评判\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>提交与评判\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>提交目标\u003C\u002Fstrong>：\u003Cul>\n\u003Cli>作业 → Fork 的个人仓库\u003C\u002Fli>\n\u003Cli>项目 → InfiniLM 主仓库\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>报告规范\u003C\u002Fstrong>：PR Comment 或 Markdown 文件，必须包含\u003Cstrong>清晰复现步骤\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw08.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kK9GglvVpJeE4xSL6vmE41EqHnI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"项目工程-infinilm-infinicore-概览\">\u003Ca class=\"anchor\" href=\"#项目工程-infinilm-infinicore-概览\" aria-label=\"跳转到标题：项目工程 InfiniLM\u002FInfiniCore 概览\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>项目工程 InfiniLM\u002FInfiniCore 概览\u003C\u002Fh2>\u003Ch3 id=\"infinicore-底层支撑\">\u003Ca class=\"anchor\" href=\"#infinicore-底层支撑\" aria-label=\"跳转到标题：InfiniCore (底层支撑)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>InfiniCore (底层支撑)\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>InfiniOP\u003C\u002Fstrong>：提供算子。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>InfiniCCL\u003C\u002Fstrong>：通信库（AllReduce, Send\u002FRecv for PP）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>InfiniRT\u003C\u002Fstrong>：运行时、CUDA Graph 支持。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>InfiniCore 封装层\u003C\u002Fstrong>：对底层算子及运行时能力进行封装，提供更简洁的调用接口，并通过缓存机制复用相关资源。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw09.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Tp%2BfnzdObCPdfvy%2B9OHOhgKw0W4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"infinilm-推理框架\">\u003Ca class=\"anchor\" href=\"#infinilm-推理框架\" aria-label=\"跳转到标题：InfiniLM (推理框架)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>InfiniLM (推理框架)\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>Pybind11\u003C\u002Fstrong>：Python\u002FC++ 绑定。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Engine\u003C\u002Fstrong>：基础引擎、分布式 Worker、边界保护。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Layers\u003C\u002Fstrong>：通用模块（Attention, Linear, MLP, Quantization）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Models\u003C\u002Fstrong>：模块化模型实现（复用 Layers，纯添加式修改）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Processor (Python)\u003C\u002Fstrong>：文本处理、EOS\u002FChatTemplate 隔离封装。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_101_zw10.jpg?Expires=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2Cdy9keEpffqPDBra7rWEViwIgM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>推理方向的征程已进入实战阶段。希望大家利用好 LLAI SYS 夯实基础，在 InfiniLM 项目中展现创造力与工程能力。预祝各位在作业与项目中取得优异成绩！\u003C\u002Fp>\n\u003C\u002Fblockquote>\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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=dUe%2BM0SLl6KCrp97rPdQhphT%2Bck%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=1792750032&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6FVLIDG%2FuTYOS25cEBcGpul%2Bzno%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,134,136,140,142,145,148,151,154,157,159,161,164,167,170,173,176,178,181,184,187,189],{"id":129,"text":130,"level":110},"关键时间节点-deadline","关键时间节点（Deadline）",{"id":132,"text":133,"level":110},"作业工程-llai-sys-解析","作业工程 LLAI SYS 解析",{"id":135,"text":135,"level":9},"项目结构",{"id":137,"text":138,"level":139},"include-核心模块","include 核心模块",4,{"id":141,"text":141,"level":110},"作业具体要求",{"id":143,"text":144,"level":9},"task-0-环境准备","Task 0: 环境准备",{"id":146,"text":147,"level":9},"task-1-张量","Task 1: 张量",{"id":149,"text":150,"level":9},"task-2-算子","Task 2: 算子",{"id":152,"text":153,"level":9},"task-3-qwen-1-5b-文本生成","Task 3: Qwen-1.5B 文本生成",{"id":155,"text":156,"level":9},"task-4-异构加速适配-重点","Task 4: 异构加速适配（重点）",{"id":158,"text":158,"level":110},"项目阶段选题指南",{"id":160,"text":160,"level":9},"选题方向推荐",{"id":162,"text":163,"level":139},"1-新模型-新结构适配-高价值","1. 新模型\u002F新结构适配（高价值）",{"id":165,"text":166,"level":139},"2-性能优化","2. 性能优化",{"id":168,"text":169,"level":139},"3-框架与功能完善","3. 框架与功能完善",{"id":171,"text":172,"level":139},"4-量化与低精度推理","4. 量化与低精度推理",{"id":174,"text":175,"level":139},"5-可靠性与工程工具","5. 可靠性与工程工具",{"id":177,"text":177,"level":9},"提交与评判",{"id":179,"text":180,"level":110},"项目工程-infinilm-infinicore-概览","项目工程 InfiniLM\u002FInfiniCore 概览",{"id":182,"text":183,"level":9},"infinicore-底层支撑","InfiniCore (底层支撑)",{"id":185,"text":186,"level":9},"infinilm-推理框架","InfiniLM (推理框架)",{"id":188,"text":188,"level":110},"加入训练营",{"id":190,"text":190,"level":110},"关注与交流"]