[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-38":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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uiV9kOffa9Xq8DT%2BbErZU0JyQ98%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8u%2BaQaeliN4QzF6MZWiB4jHYDCE%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8pXXbnSlNEtVJCJ1e4RzUlqmRn8%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_38_fm.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qe3kGGVuIfB1ast9JXbJKuukzRE%3D","InfiniTensor,冬季训练营,大模型推理,分布式推理,计算优化","讲解大模型计算优化与分布式推理技术，覆盖并行策略、吞吐提升与部署关键点。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营推理引擎第四课《大模型计算优化与分布式推理》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ccPD%2FluTZJFb7tUeG9nIapdGsv0%3D)\r\n\r\n2026 年 1 月 26 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎系列第四课《大模型计算优化与分布式推理》开讲。\r\n\r\n本节课介绍计算优化与分布式推理技术，充分利用硬件资源，降低推理成本，提供实时高效服务。\r\n\r\n### 背景与意义\r\n#### 1.Scaling Law 驱动模型容量越来越大：\r\n- Scaling Law 显示，在一定范围内，增加模型参数量、训练数据量和计算量，模型性能通常会按可预测规律提升。这使得模型参数量越来越大。\r\n#### 2. 大模型带来的问题：\r\n- 空间占用高、计算延迟、KV Cache 所需空间更大、硬件成本高、API 价格昂贵。\r\n\r\n### 低精度类型\r\n#### 1. 低精度数据类型优势：\r\nFP32、FP16、BF16、FP8等，不同数据类型占用的位数和表示范围不同。降低模型文件和内存显存占用，例如 8B 模型从 FP32 降到 FP16，可以显著减少显存占用。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2B2P9JDfWGrT225MYNmqI3%2FZF44%3D)\r\n\r\n#### 2. 量化推理：\r\n- **定义：** 牺牲计算精度换取更快计算速度和更小内存占用。\r\n- **分类：** 训练时量化（训练过程中考虑权重量化）和训练后量化（PTQ，训练结束后通过量化算法处理权重文件）。\r\n- **量化过程：** 浮点数运算映射到低精度整数运算，如 FP32 映射到 INT8。\r\n- **量化优点：** 权重文件变小，访存加快，INT8 有专门核心加速计算，降低部署成本。\r\n- **量化缺点：** 精度损失，模型性能变差。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lFBSSjDduksSIl135zHORAY183Y%3D)\r\n\r\n\r\n### 融合算子\r\n#### 1. 定义：\r\n- 将多个算子合并成一个算子，减少内存访问和 kernel 发射次数。\r\n#### 2. 常见融合场景：\r\n- 矩阵乘法融合\r\n- MLP 模块融合\r\n- Decoder 模块\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LzLxyQU3yNbmYyH3X9WKrE6SczM%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wiymi1Jei1TfrbYdlwZAvhdZL%2Fk%3D)\r\n\r\n\r\n\r\n### 模型结构\r\n#### 1. 目标：提高计算效率，降低硬件要求。\r\n#### 2. 注意力计算优化：\r\n- **MHA 到 MQA 到 GQA：** 平衡显存占用和模型性能。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7tdYSjGEvaPngopNHHfper8KKqU%3D)\r\n\r\n- **Flash Attention：** 结合 GPU 硬件特性，优化注意力计算，减少显存占用，提高计算速度。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=t4UtVIvkXLSRrvKQP3uBfjPj9lI%3D)\r\n\r\n\r\n#### 3. KV Cache 管理优化：\r\n- 传统方法问题：内存碎片化、重复计算。\r\n- Paged Attention：借鉴操作系统虚拟内存分页思想，按需分配显存，减少内存碎片化。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KHFdFhLw%2FNNGke5M3NsYCaqbfTA%3D)\r\n\r\n\r\n### 分布式推理\r\n#### 1. 定义：\r\n- 通过多张显卡或多个计算节点共同完成推理任务。\r\n\r\n#### 2. 并行策略：\r\n- **数据并行：** 每个机器拥有完整权重参数，输入根据调度分配到不同机器，提高吞吐量。\r\n- **模型并行：** 每个机器只拥有部分权重参数，多张卡共同计算，通过通信汇总结果。\r\n- **流水并行：** 模型切分为多个阶段，不同 GPU 负责不同阶段计算。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mphxZsGmQOEaaMBUIsNGNOvL4K0%3D)\r\n\r\n#### 3. 张量切分方式：\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw10.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=V1WNK9LPSGl3HQTYtnykMFZDeEY%3D)\r\n\r\n- 按行切分：矩阵横着切，上半部分在一张卡上，下半部分在另一张卡上。\r\n- 按列切分：矩阵竖着切，左边在一张卡上，右边在另一张卡上。\r\n\r\n- 矩阵分块相乘：将大矩阵拆分为小矩阵进行计算，结果相同。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw11.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YxDtugnclJP4U1hOimUGDEMed7g%3D)\r\n\r\n\r\n\r\n#### 4. 具体实现：\r\n- MLP 模块权重切分：up、gate 权重竖切，down 权重横切，通过 AllReduce 通信得到结果。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw12.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tI0EGWR8fvvZb35r8ZzZnphVsbE%3D)\r\n\r\n- Attention 模块权重切分：Q、K、V 权重切第二个维度，O 权重切第零个维度，过程与 MLP 切分相同。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw13.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fk3LTN5FMrHA8sRFDxVupiEy3bA%3D)\r\n\r\n- Embedding 切分：token 转向量过程竖着切，通过 AllGather 通信。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw14.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=i9y1WtHSdfXyWOS0Lk8nceFbbEc%3D)\r\n\r\n### 总结\r\n本节课主要从低精度类型、融合算子、模型结构以及分布式四个方面讲解了大模型计算优化与分布式推理技术。后续推理课程将涉及大模型服务系统。\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_01_zw6.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DEth0OhERyx68xufLNiv%2BIt7foE%3D)\r\n\r\n### 💡 学习贴士\r\n更多细节及完整课程内容，请观看直播或查看课程回放。\r\n\r\n链接：https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\r\n\r\n### 📚 训练营报名持续进行中\r\n扫码报名与加入群聊，或点击 **「阅读原文」** 直达官网报名\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw7.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OXZH4IYHRMHbFPG7EYCBYHhSMto%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-37","study-37","InfiniTensor 冬季训练营训练系统第三课《大模型训练中的显存优化》",71,"\u002Fstudy-material\u002Fstudy-38","study-38","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-39","study-39","InfiniTensor 冬季训练营 Triton&九齿第五课《天数智芯 Triton 实战》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ccPD%2FluTZJFb7tUeG9nIapdGsv0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 26 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎系列第四课《大模型计算优化与分布式推理》开讲。\u003C\u002Fp>\n\u003Cp>本节课介绍计算优化与分布式推理技术，充分利用硬件资源，降低推理成本，提供实时高效服务。\u003C\u002Fp>\n\u003Ch3 id=\"背景与意义\">\u003Ca class=\"anchor\" href=\"#背景与意义\" aria-label=\"跳转到标题：背景与意义\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>背景与意义\u003C\u002Fh3>\u003Ch4 id=\"1-scaling-law-驱动模型容量越来越大\">\u003Ca class=\"anchor\" href=\"#1-scaling-law-驱动模型容量越来越大\" aria-label=\"跳转到标题：1.Scaling Law 驱动模型容量越来越大：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.Scaling Law 驱动模型容量越来越大：\u003C\u002Fh4>\u003Cul>\n\u003Cli>Scaling Law 显示，在一定范围内，增加模型参数量、训练数据量和计算量，模型性能通常会按可预测规律提升。这使得模型参数量越来越大。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-大模型带来的问题\">\u003Ca class=\"anchor\" href=\"#2-大模型带来的问题\" aria-label=\"跳转到标题：2. 大模型带来的问题：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 大模型带来的问题：\u003C\u002Fh4>\u003Cul>\n\u003Cli>空间占用高、计算延迟、KV Cache 所需空间更大、硬件成本高、API 价格昂贵。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"低精度类型\">\u003Ca class=\"anchor\" href=\"#低精度类型\" aria-label=\"跳转到标题：低精度类型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>低精度类型\u003C\u002Fh3>\u003Ch4 id=\"1-低精度数据类型优势\">\u003Ca class=\"anchor\" href=\"#1-低精度数据类型优势\" aria-label=\"跳转到标题：1. 低精度数据类型优势：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 低精度数据类型优势：\u003C\u002Fh4>\u003Cp>FP32、FP16、BF16、FP8等，不同数据类型占用的位数和表示范围不同。降低模型文件和内存显存占用，例如 8B 模型从 FP32 降到 FP16，可以显著减少显存占用。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n%2B2P9JDfWGrT225MYNmqI3%2FZF44%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-量化推理\">\u003Ca class=\"anchor\" href=\"#2-量化推理\" aria-label=\"跳转到标题：2. 量化推理：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 量化推理：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>定义：\u003C\u002Fstrong> 牺牲计算精度换取更快计算速度和更小内存占用。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>分类：\u003C\u002Fstrong> 训练时量化（训练过程中考虑权重量化）和训练后量化（PTQ，训练结束后通过量化算法处理权重文件）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>量化过程：\u003C\u002Fstrong> 浮点数运算映射到低精度整数运算，如 FP32 映射到 INT8。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>量化优点：\u003C\u002Fstrong> 权重文件变小，访存加快，INT8 有专门核心加速计算，降低部署成本。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>量化缺点：\u003C\u002Fstrong> 精度损失，模型性能变差。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lFBSSjDduksSIl135zHORAY183Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"融合算子\">\u003Ca class=\"anchor\" href=\"#融合算子\" aria-label=\"跳转到标题：融合算子\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>融合算子\u003C\u002Fh3>\u003Ch4 id=\"1-定义\">\u003Ca class=\"anchor\" href=\"#1-定义\" aria-label=\"跳转到标题：1. 定义：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 定义：\u003C\u002Fh4>\u003Cul>\n\u003Cli>将多个算子合并成一个算子，减少内存访问和 kernel 发射次数。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-常见融合场景\">\u003Ca class=\"anchor\" href=\"#2-常见融合场景\" aria-label=\"跳转到标题：2. 常见融合场景：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 常见融合场景：\u003C\u002Fh4>\u003Cul>\n\u003Cli>矩阵乘法融合\u003C\u002Fli>\n\u003Cli>MLP 模块融合\u003C\u002Fli>\n\u003Cli>Decoder 模块\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LzLxyQU3yNbmYyH3X9WKrE6SczM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wiymi1Jei1TfrbYdlwZAvhdZL%2Fk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"模型结构\">\u003Ca class=\"anchor\" href=\"#模型结构\" aria-label=\"跳转到标题：模型结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型结构\u003C\u002Fh3>\u003Ch4 id=\"1-目标-提高计算效率-降低硬件要求\">\u003Ca class=\"anchor\" href=\"#1-目标-提高计算效率-降低硬件要求\" aria-label=\"跳转到标题：1. 目标：提高计算效率，降低硬件要求。\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 目标：提高计算效率，降低硬件要求。\u003C\u002Fh4>\u003Ch4 id=\"2-注意力计算优化\">\u003Ca class=\"anchor\" href=\"#2-注意力计算优化\" aria-label=\"跳转到标题：2. 注意力计算优化：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 注意力计算优化：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>MHA 到 MQA 到 GQA：\u003C\u002Fstrong> 平衡显存占用和模型性能。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7tdYSjGEvaPngopNHHfper8KKqU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Flash Attention：\u003C\u002Fstrong> 结合 GPU 硬件特性，优化注意力计算，减少显存占用，提高计算速度。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=t4UtVIvkXLSRrvKQP3uBfjPj9lI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-kv-cache-管理优化\">\u003Ca class=\"anchor\" href=\"#3-kv-cache-管理优化\" aria-label=\"跳转到标题：3. KV Cache 管理优化：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. KV Cache 管理优化：\u003C\u002Fh4>\u003Cul>\n\u003Cli>传统方法问题：内存碎片化、重复计算。\u003C\u002Fli>\n\u003Cli>Paged Attention：借鉴操作系统虚拟内存分页思想，按需分配显存，减少内存碎片化。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KHFdFhLw%2FNNGke5M3NsYCaqbfTA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"分布式推理\">\u003Ca class=\"anchor\" href=\"#分布式推理\" aria-label=\"跳转到标题：分布式推理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>分布式推理\u003C\u002Fh3>\u003Ch4 id=\"1-定义-1\">\u003Ca class=\"anchor\" href=\"#1-定义-1\" aria-label=\"跳转到标题：1. 定义：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 定义：\u003C\u002Fh4>\u003Cul>\n\u003Cli>通过多张显卡或多个计算节点共同完成推理任务。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-并行策略\">\u003Ca class=\"anchor\" href=\"#2-并行策略\" aria-label=\"跳转到标题：2. 并行策略：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 并行策略：\u003C\u002Fh4>\u003Cul>\n\u003Cli>\u003Cstrong>数据并行：\u003C\u002Fstrong> 每个机器拥有完整权重参数，输入根据调度分配到不同机器，提高吞吐量。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>模型并行：\u003C\u002Fstrong> 每个机器只拥有部分权重参数，多张卡共同计算，通过通信汇总结果。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>流水并行：\u003C\u002Fstrong> 模型切分为多个阶段，不同 GPU 负责不同阶段计算。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mphxZsGmQOEaaMBUIsNGNOvL4K0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-张量切分方式\">\u003Ca class=\"anchor\" href=\"#3-张量切分方式\" aria-label=\"跳转到标题：3. 张量切分方式：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 张量切分方式：\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw10.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=V1WNK9LPSGl3HQTYtnykMFZDeEY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cp>按行切分：矩阵横着切，上半部分在一张卡上，下半部分在另一张卡上。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>按列切分：矩阵竖着切，左边在一张卡上，右边在另一张卡上。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>矩阵分块相乘：将大矩阵拆分为小矩阵进行计算，结果相同。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw11.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YxDtugnclJP4U1hOimUGDEMed7g%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"4-具体实现\">\u003Ca class=\"anchor\" href=\"#4-具体实现\" aria-label=\"跳转到标题：4. 具体实现：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. 具体实现：\u003C\u002Fh4>\u003Cul>\n\u003Cli>MLP 模块权重切分：up、gate 权重竖切，down 权重横切，通过 AllReduce 通信得到结果。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw12.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tI0EGWR8fvvZb35r8ZzZnphVsbE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Attention 模块权重切分：Q、K、V 权重切第二个维度，O 权重切第零个维度，过程与 MLP 切分相同。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw13.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fk3LTN5FMrHA8sRFDxVupiEy3bA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Embedding 切分：token 转向量过程竖着切，通过 AllGather 通信。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_38_zw14.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=i9y1WtHSdfXyWOS0Lk8nceFbbEc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 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\u002Fh3>\u003Cp>本节课主要从低精度类型、融合算子、模型结构以及分布式四个方面讲解了大模型计算优化与分布式推理技术。后续推理课程将涉及大模型服务系统。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DEth0OhERyx68xufLNiv%2BIt7foE%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_01_zw7.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OXZH4IYHRMHbFPG7EYCBYHhSMto%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,134,137,139,142,145,147,150,153,155,158,161,164,166,168,171,174,177,179,182],{"id":129,"text":129,"level":9},"背景与意义",{"id":131,"text":132,"level":133},"1-scaling-law-驱动模型容量越来越大","1.Scaling Law 驱动模型容量越来越大：",4,{"id":135,"text":136,"level":133},"2-大模型带来的问题","2. 大模型带来的问题：",{"id":138,"text":138,"level":9},"低精度类型",{"id":140,"text":141,"level":133},"1-低精度数据类型优势","1. 低精度数据类型优势：",{"id":143,"text":144,"level":133},"2-量化推理","2. 量化推理：",{"id":146,"text":146,"level":9},"融合算子",{"id":148,"text":149,"level":133},"1-定义","1. 定义：",{"id":151,"text":152,"level":133},"2-常见融合场景","2. 常见融合场景：",{"id":154,"text":154,"level":9},"模型结构",{"id":156,"text":157,"level":133},"1-目标-提高计算效率-降低硬件要求","1. 目标：提高计算效率，降低硬件要求。",{"id":159,"text":160,"level":133},"2-注意力计算优化","2. 注意力计算优化：",{"id":162,"text":163,"level":133},"3-kv-cache-管理优化","3. KV Cache 管理优化：",{"id":165,"text":165,"level":9},"分布式推理",{"id":167,"text":149,"level":133},"1-定义-1",{"id":169,"text":170,"level":133},"2-并行策略","2. 并行策略：",{"id":172,"text":173,"level":133},"3-张量切分方式","3. 张量切分方式：",{"id":175,"text":176,"level":133},"4-具体实现","4. 具体实现：",{"id":178,"text":178,"level":9},"总结",{"id":180,"text":181,"level":9},"学习贴士","💡 学习贴士",{"id":183,"text":184,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]