[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-15":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=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tqXtkjDVEuQBnsL07LARCfQJvdE%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=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tM28gi%2FKPywJvel7AqVcxFGnKTg%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=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RSwSsL0Rw0AFxM%2Bz65pGHzs2YLw%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_15_fm.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tCxQW03WbUwiMxG01UtVWu8covc%3D","InfiniTensor,冬季训练营,大模型原理,模型结构,推理引擎","讲解大模型原理与结构基础，梳理核心组件、推理流程与模型能力形成机制。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营推理引擎第二课《大模型原理与结构》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw01.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TP%2BSXD0nIW%2FBIio7lAS%2Be7uo%2FTk%3D)\r\n\r\n\r\n\r\n2026 年 1 月 14 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎方向第二课《大模型原理与结构》开讲。\r\n\r\n本节课正式开启大模型推理系统的核心内容讲解，作为六节系列课程的第二课，重点解析支撑现代大语言模型的 Transformer 架构与自注意力机制。\r\n\r\n\r\n### 人工神经网络基础\r\n#### 1. 向量到向量（Vector to Vector）任务\r\n- 输入输出均为向量，如手写数字识别。\r\n- 局限：无法处理向量间关系，如序列任务。\r\n#### 2. 序列到序列（Sequence to Sequence）任务\r\n- 输入输出为向量序列，如翻译、对话、文本生成。\r\n- 特点：输入输出向量数量任意，且输入向量间相互影响。\r\n#### 3. 词性标注任务示例\r\n- 输入文字，输出每个词的词性。\r\n- 感知器无法处理上下文信息，引出上下文处理需求。\r\n\r\n### 上下文信息处理\r\n#### 1. 滑动窗口方法\r\n- 原理：将上下文绑定输入模型。\r\n- 缺点：窗口大小固定，无法捕捉长距离上下文。\r\n#### 2. 循环神经网络（RNN）\r\n- 原理：维护记忆模块，随输入序列更新。\r\n- 缺点：长链条信息弱化，计算顺序性强，降低效率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw02.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KSiEZxQ%2BFFTya85E622tkC5HIPI%3D)\r\n\r\n#### 3. 注意力机制（Attention Mechanism）\r\n- 上下文信息通过计算输入之间的相似度分数，并经 Softmax 归一化得到注意力权重分布\r\n- 计算流程：query、key、value 向量映射，计算相关程度，Softmax 处理，得到注意力输出。\r\n- 具体计算流程：它通过将输入的词向量（如“道”）与学习到的权重矩阵相乘，分别映射为 Query（查询）、**Key（键）**和 Value（值） 三个向量，并通过计算 Query 与 Key 的点积（如图中的 $a = q cdot k = -1.1700$）来衡量不同输入之间的相关性权重，从而实现模型对重要特征的聚焦。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw03.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wkKqzGwUEbhogItEP%2FSV7NdUNlI%3D)\r\n\r\n- Softmax 的原理：它通过指数映射和归一化处理，将任意范围的激活值转换为总和为 1 的概率分布；\r\n- 数值稳定版本 Safe Softmax：为了防止在计算过程中因输入值过大导致数值溢出，Safe Softmax 引入了减去向量最大值 ($m = max(x_i)$) 的技巧，在保证数学等价性的前提下提升了算子在工程实现中的稳定性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw04.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kdFKbfnVazNjrw90NCkgQjwY8dQ%3D)\r\n\r\n\r\n\r\n### 自注意力机制\r\n#### 1. 自注意力层计算\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw05.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UMqHlVKL0jwZ0GZZMNKUD7sCWyE%3D)\r\n\r\n#### 2. 位置信息处理\r\n- 旋转位置编码（Rotary Position Embedding, RoPE）：通过角度表示位置关系，保持计算稳定性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw06.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rAQEEiOBiihz7y5jI%2BoEV0cbagU%3D)\r\n\r\n- 主流算法：GPT-J 与 GPT-Neox，数学无本质区别，角度排布不同。\r\n#### 3. 多头注意力机制\r\n- 原理：通过将 q、k、v 向量在特征维度上拆分为多个子空间（heads），并行建模不同子空间的注意力关系。\r\n- 优势：并行计算，提高效率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw07.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0G%2BgoLcXyGS2WA9%2Bta5daGE040s%3D)\r\n\r\n\r\n### 大模型结构详解\r\n#### 1. Decoder 模块\r\n- 组成：自注意力层与 MLP 层首尾相连。\r\n- 层归一化（RMSNorm）与残差结构：保证计算稳定，防止信息损失。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw08.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=olu%2FKf9D0fl%2Brmy0zNsEM9JXmpU%3D)\r\n\r\n\r\n#### 2. MLP 层结构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw09.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j6hnY7%2FElGy8TOucm%2F9cY2x03ks%3D)\r\n\r\n- 参数：hidden size 与 intermediate size，控制输入输出维度与中间层维度。\r\n\r\n### 文本向量化处理\r\n#### 1. 词表构建与分词方法\r\n- 词表构建：BPE 算法，高频组合优先成词。\r\n- Tokenizer 作用：将文本转换为 toke-id 序列。\r\n#### 2. 向量表示方法\r\n- One-hot 编码：简单但低效。\r\n- 输入嵌入层（Input Embedding）：神经网络映射 token-id 为连续向量，保留语义信息。\r\n- 输出嵌入层（Output Embedding）：将向量映射回词表长度概率分布，用于文本生成。\r\n\r\n\r\n## 本专业阶段作业\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw10.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CSG6WnUJU1DQf9mrHGiJeew0Lp8%3D)\r\n\r\n\r\n\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=epXM94790JRyVZK5KPiIf95EZZQ%3D)\r\n\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=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=T6W1BlHdoBa2bd0P%2B0OABUI5sTc%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-14","study-14","InfiniTensor 冬季训练营通信与并行第一课《互联、集群与通信》",94,"\u002Fstudy-material\u002Fstudy-15","study-15","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-16","study-16","InfiniTensor 冬季训练营前沿模型第二课《科学多模态大模型》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw01.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TP%2BSXD0nIW%2FBIio7lAS%2Be7uo%2FTk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 14 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营推理引擎方向第二课《大模型原理与结构》开讲。\u003C\u002Fp>\n\u003Cp>本节课正式开启大模型推理系统的核心内容讲解，作为六节系列课程的第二课，重点解析支撑现代大语言模型的 Transformer 架构与自注意力机制。\u003C\u002Fp>\n\u003Ch3 id=\"人工神经网络基础\">\u003Ca class=\"anchor\" href=\"#人工神经网络基础\" aria-label=\"跳转到标题：人工神经网络基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>人工神经网络基础\u003C\u002Fh3>\u003Ch4 id=\"1-向量到向量-vector-to-vector-任务\">\u003Ca class=\"anchor\" href=\"#1-向量到向量-vector-to-vector-任务\" aria-label=\"跳转到标题：1. 向量到向量（Vector to Vector）任务\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 向量到向量（Vector to Vector）任务\u003C\u002Fh4>\u003Cul>\n\u003Cli>输入输出均为向量，如手写数字识别。\u003C\u002Fli>\n\u003Cli>局限：无法处理向量间关系，如序列任务。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-序列到序列-sequence-to-sequence-任务\">\u003Ca class=\"anchor\" href=\"#2-序列到序列-sequence-to-sequence-任务\" aria-label=\"跳转到标题：2. 序列到序列（Sequence to Sequence）任务\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 序列到序列（Sequence to Sequence）任务\u003C\u002Fh4>\u003Cul>\n\u003Cli>输入输出为向量序列，如翻译、对话、文本生成。\u003C\u002Fli>\n\u003Cli>特点：输入输出向量数量任意，且输入向量间相互影响。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-词性标注任务示例\">\u003Ca class=\"anchor\" href=\"#3-词性标注任务示例\" aria-label=\"跳转到标题：3. 词性标注任务示例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 词性标注任务示例\u003C\u002Fh4>\u003Cul>\n\u003Cli>输入文字，输出每个词的词性。\u003C\u002Fli>\n\u003Cli>感知器无法处理上下文信息，引出上下文处理需求。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"上下文信息处理\">\u003Ca class=\"anchor\" href=\"#上下文信息处理\" aria-label=\"跳转到标题：上下文信息处理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>上下文信息处理\u003C\u002Fh3>\u003Ch4 id=\"1-滑动窗口方法\">\u003Ca class=\"anchor\" href=\"#1-滑动窗口方法\" aria-label=\"跳转到标题：1. 滑动窗口方法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 滑动窗口方法\u003C\u002Fh4>\u003Cul>\n\u003Cli>原理：将上下文绑定输入模型。\u003C\u002Fli>\n\u003Cli>缺点：窗口大小固定，无法捕捉长距离上下文。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-循环神经网络-rnn\">\u003Ca class=\"anchor\" href=\"#2-循环神经网络-rnn\" aria-label=\"跳转到标题：2. 循环神经网络（RNN）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 循环神经网络（RNN）\u003C\u002Fh4>\u003Cul>\n\u003Cli>原理：维护记忆模块，随输入序列更新。\u003C\u002Fli>\n\u003Cli>缺点：长链条信息弱化，计算顺序性强，降低效率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw02.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KSiEZxQ%2BFFTya85E622tkC5HIPI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-注意力机制-attention-mechanism\">\u003Ca class=\"anchor\" href=\"#3-注意力机制-attention-mechanism\" aria-label=\"跳转到标题：3. 注意力机制（Attention Mechanism）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 注意力机制（Attention Mechanism）\u003C\u002Fh4>\u003Cul>\n\u003Cli>上下文信息通过计算输入之间的相似度分数，并经 Softmax 归一化得到注意力权重分布\u003C\u002Fli>\n\u003Cli>计算流程：query、key、value 向量映射，计算相关程度，Softmax 处理，得到注意力输出。\u003C\u002Fli>\n\u003Cli>具体计算流程：它通过将输入的词向量（如“道”）与学习到的权重矩阵相乘，分别映射为 Query（查询）、**Key（键）**和 Value（值） 三个向量，并通过计算 Query 与 Key 的点积（如图中的 $a = q cdot k = -1.1700$）来衡量不同输入之间的相关性权重，从而实现模型对重要特征的聚焦。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw03.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wkKqzGwUEbhogItEP%2FSV7NdUNlI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Softmax 的原理：它通过指数映射和归一化处理，将任意范围的激活值转换为总和为 1 的概率分布；\u003C\u002Fli>\n\u003Cli>数值稳定版本 Safe Softmax：为了防止在计算过程中因输入值过大导致数值溢出，Safe Softmax 引入了减去向量最大值 ($m = max(x_i)$) 的技巧，在保证数学等价性的前提下提升了算子在工程实现中的稳定性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw04.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kdFKbfnVazNjrw90NCkgQjwY8dQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"自注意力机制\">\u003Ca class=\"anchor\" href=\"#自注意力机制\" aria-label=\"跳转到标题：自注意力机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>自注意力机制\u003C\u002Fh3>\u003Ch4 id=\"1-自注意力层计算\">\u003Ca class=\"anchor\" href=\"#1-自注意力层计算\" aria-label=\"跳转到标题：1. 自注意力层计算\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 自注意力层计算\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw05.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UMqHlVKL0jwZ0GZZMNKUD7sCWyE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-位置信息处理\">\u003Ca class=\"anchor\" href=\"#2-位置信息处理\" aria-label=\"跳转到标题：2. 位置信息处理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 位置信息处理\u003C\u002Fh4>\u003Cul>\n\u003Cli>旋转位置编码（Rotary Position Embedding, RoPE）：通过角度表示位置关系，保持计算稳定性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw06.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rAQEEiOBiihz7y5jI%2BoEV0cbagU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>主流算法：GPT-J 与 GPT-Neox，数学无本质区别，角度排布不同。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-多头注意力机制\">\u003Ca class=\"anchor\" href=\"#3-多头注意力机制\" aria-label=\"跳转到标题：3. 多头注意力机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 多头注意力机制\u003C\u002Fh4>\u003Cul>\n\u003Cli>原理：通过将 q、k、v 向量在特征维度上拆分为多个子空间（heads），并行建模不同子空间的注意力关系。\u003C\u002Fli>\n\u003Cli>优势：并行计算，提高效率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw07.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0G%2BgoLcXyGS2WA9%2Bta5daGE040s%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"大模型结构详解\">\u003Ca class=\"anchor\" href=\"#大模型结构详解\" aria-label=\"跳转到标题：大模型结构详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>大模型结构详解\u003C\u002Fh3>\u003Ch4 id=\"1-decoder-模块\">\u003Ca class=\"anchor\" href=\"#1-decoder-模块\" aria-label=\"跳转到标题：1. Decoder 模块\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. Decoder 模块\u003C\u002Fh4>\u003Cul>\n\u003Cli>组成：自注意力层与 MLP 层首尾相连。\u003C\u002Fli>\n\u003Cli>层归一化（RMSNorm）与残差结构：保证计算稳定，防止信息损失。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw08.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=olu%2FKf9D0fl%2Brmy0zNsEM9JXmpU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-mlp-层结构\">\u003Ca class=\"anchor\" href=\"#2-mlp-层结构\" aria-label=\"跳转到标题：2. MLP 层结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. MLP 层结构\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw09.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j6hnY7%2FElGy8TOucm%2F9cY2x03ks%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>参数：hidden size 与 intermediate size，控制输入输出维度与中间层维度。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"文本向量化处理\">\u003Ca class=\"anchor\" href=\"#文本向量化处理\" aria-label=\"跳转到标题：文本向量化处理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>文本向量化处理\u003C\u002Fh3>\u003Ch4 id=\"1-词表构建与分词方法\">\u003Ca class=\"anchor\" href=\"#1-词表构建与分词方法\" aria-label=\"跳转到标题：1. 词表构建与分词方法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 词表构建与分词方法\u003C\u002Fh4>\u003Cul>\n\u003Cli>词表构建：BPE 算法，高频组合优先成词。\u003C\u002Fli>\n\u003Cli>Tokenizer 作用：将文本转换为 toke-id 序列。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-向量表示方法\">\u003Ca class=\"anchor\" href=\"#2-向量表示方法\" aria-label=\"跳转到标题：2. 向量表示方法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 向量表示方法\u003C\u002Fh4>\u003Cul>\n\u003Cli>One-hot 编码：简单但低效。\u003C\u002Fli>\n\u003Cli>输入嵌入层（Input Embedding）：神经网络映射 token-id 为连续向量，保留语义信息。\u003C\u002Fli>\n\u003Cli>输出嵌入层（Output Embedding）：将向量映射回词表长度概率分布，用于文本生成。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"本专业阶段作业\">\u003Ca class=\"anchor\" href=\"#本专业阶段作业\" aria-label=\"跳转到标题：本专业阶段作业\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>本专业阶段作业\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_15_zw10.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CSG6WnUJU1DQf9mrHGiJeew0Lp8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_01_zw6.jpg?Expires=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=epXM94790JRyVZK5KPiIf95EZZQ%3D\" alt=\"\">\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=1792752826&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=T6W1BlHdoBa2bd0P%2B0OABUI5sTc%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,134,137,140,142,145,148,151,153,156,159,162,164,167,170,172,175,178,180,183],{"id":129,"text":129,"level":9},"人工神经网络基础",{"id":131,"text":132,"level":133},"1-向量到向量-vector-to-vector-任务","1. 向量到向量（Vector to Vector）任务",4,{"id":135,"text":136,"level":133},"2-序列到序列-sequence-to-sequence-任务","2. 序列到序列（Sequence to Sequence）任务",{"id":138,"text":139,"level":133},"3-词性标注任务示例","3. 词性标注任务示例",{"id":141,"text":141,"level":9},"上下文信息处理",{"id":143,"text":144,"level":133},"1-滑动窗口方法","1. 滑动窗口方法",{"id":146,"text":147,"level":133},"2-循环神经网络-rnn","2. 循环神经网络（RNN）",{"id":149,"text":150,"level":133},"3-注意力机制-attention-mechanism","3. 注意力机制（Attention Mechanism）",{"id":152,"text":152,"level":9},"自注意力机制",{"id":154,"text":155,"level":133},"1-自注意力层计算","1. 自注意力层计算",{"id":157,"text":158,"level":133},"2-位置信息处理","2. 位置信息处理",{"id":160,"text":161,"level":133},"3-多头注意力机制","3. 多头注意力机制",{"id":163,"text":163,"level":9},"大模型结构详解",{"id":165,"text":166,"level":133},"1-decoder-模块","1. Decoder 模块",{"id":168,"text":169,"level":133},"2-mlp-层结构","2. MLP 层结构",{"id":171,"text":171,"level":9},"文本向量化处理",{"id":173,"text":174,"level":133},"1-词表构建与分词方法","1. 词表构建与分词方法",{"id":176,"text":177,"level":133},"2-向量表示方法","2. 向量表示方法",{"id":179,"text":179,"level":110},"本专业阶段作业",{"id":181,"text":182,"level":9},"学习贴士","💡 学习贴士",{"id":184,"text":185,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]