[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-64":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_64_fm.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Z0Jm5hCPjMxRh819tjdFPgXZK7I%3D","Transformer,自注意力,RoPE,大模型 Decoder,Tokenizer","课程拆解 Transformer 架构，对比 RNN 缺陷，讲解自注意力 QKV、RoPE 位置编码、多头注意力、Decoder 组件，介绍分词、Embedding 等大模型基础模块。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营推理引擎 2：《大模型原理与结构》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Z0gO%2FAENry9dvhWZT58a9cXI3oc%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 16 日，2026 夏季 InfiniTensor 训练营推理引擎方向第二课《大模型原理与结构》开讲。\r\n推理引擎系列课程共 6 节，前置课程《AI 模型导论与推理基础原理》介绍了人工神经网络的基础知识，vec2vec 任务、矩阵乘、激活函数等内容，从本节课开始，正式进入大模型推理系统的核心领域。作为系列课程的第二讲，我们将深入拆解大模型的基石——Transformer 架构。\u003C\u002Fspan>\r\n\r\n\r\n## 为什么需要 Transformer？\r\n\r\n### 传统神经网络的局限\r\n1. **Vector to Vector**：输入输出维度固定，不便于直接处理变长序列（如手写数字识别）。\r\n2. **独立计算缺陷**：矩阵乘法中各输入向量相互独立，无法感知“道可道”中不同位置“道”字的语义差异。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TQ4aQNI4Q99qEY0Jck6l%2BrxIdRM%3D)\r\n\r\n### Sequence to Sequence\r\n1. **变长映射**：输入和输出均可为变长序列，二者长度不必相等（如翻译、文本生成、语音识别）。\r\n2. **上下文依赖**：词义随位置变化，需捕捉全局信息。\r\n\r\n### 早期解决方案及其瓶颈\r\n1. **滑动窗口**：仅能捕获局部固定范围上下文。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LMv%2BYEXo50E8el3Ote9TWOH8PUQ%3D)\r\n\r\n2. **循环神经网络 (RNN)**：\r\n  - 优点：理论支持无限长上下文（记忆模块 H 传递）。\r\n  - 缺点：梯度消失\u002F信息衰减；串行计算导致硬件算力利用率低，无法并行。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FsnK85tDhxR%2BDM8D8NK29uuTvBI%3D)\r\n\r\n\r\n## 自注意力机制 (Self-Attention)\r\n\r\n### 核心思想\r\n- 上下文 = 当前输入与其他所有输入的相关性分布。\r\n- 突破窗口限制，实现全局信息交互。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=waAEgPSbUEDE3QgZsw9EP4W7ClM%3D)\r\n\r\n### `Q`-`K`-`V` 计算流程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mZ3Xq%2FUYcBaJdvp%2Fv3nZHdU4DuI%3D)\r\n\r\n- **线性映射**：通过神经网络计算得到 `Query`、`Key`、`Value` 矩阵。\r\n- **注意力分数**：$alpha = q cdot k$\r\n- **大模型算子 `Softmax`**：将分数转换为概率分布（和为 1）。\r\n- **`Safe Softmax`**：减去最大值防止指数爆炸，保证数值稳定性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4yA2OtQ6f52qiFe6Y3U%2F3GkroyI%3D)\r\n\r\n\r\n- **因果掩码 (`Causal Mask`)**：主要用于自回归注意力，屏蔽未来 Token，防止训练时“作弊”。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sqbOmpclFeAC92TiFcxppW8u0K0%3D)\r\n\r\n\r\n- **矩阵并行化**：将逐对向量运算转化为矩阵乘法，充分利用 GPU 算力。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CfqDAZ9XPmGGctojxBHUkVrBIc8%3D)\r\n\r\n\r\n### 旋转位置编码 (`RoPE`)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw10.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=q28hWd33PeYJ2ZEHsKAAtEzbKuE%3D)\r\n\r\n\r\n- **解决痛点**：不包含位置信息的自注意力具有排列等变性，本身无法区分 Token 的先后顺序，因此需要显式引入位置信息。\r\n- **原理**：通过绝对位置的三角函数（`sin`\u002F`cos`）对 `Q`、`K` 向量进行几何旋转。\r\n- **优势**：相对位置信息自然融入点积运算；角度表可预计算，推理高效。\r\n\r\n### 自注意力层\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw11.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QWuz5kIToJ8kF0N3wdJzI7fzCx0%3D)\r\n\r\n\r\n### 多头注意力\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw12.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fLOZhG0g8YSSxDL%2BNKgrSoCmIrk%3D)\r\n\r\n- **动机**：隐藏表示划分为多个注意力头，使模型能够在不同表示子空间中并行学习不同类型的语义关系和位置关系。\r\n- **机制**：多组独立的 `Q`、`K`、`V` 并行计算，捕获不同子空间的语义相关性。\r\n\r\n\r\n## 解码器（Decoder）模块\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw13.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5oszdZoOeOzUYOVf3M9o2NQuu7E%3D)\r\n\r\n### Decoder 模块标准组件\r\n- **`RMSNorm`**：基于均方根的归一化方法，通过对隐藏状态进行尺度归一化，改善深层网络训练和推理过程中的数值稳定性。\r\n- **`Residual`**：残差连接，缓解梯度消失，保留原始信息。\r\n- **多层感知器 `MLP`**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw14.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HgJD9BYcIghd9%2BfsjsouN7W5UiQ%3D)\r\n\r\n\r\n### 如何构建词表\r\n1. **分词算法**：BPE（字节对编码）、WordPiece、Unigram；平衡词表大小与语义完整性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw15.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=om42t9Lmvn%2F0snqHTI7VunU%2B5ro%3D)\r\n\r\n2. **`Tokenizer`**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw16.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=H%2Bg1fbq2%2FDh4LwJLlZyidwzbHQQ%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw17.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=v3yOd%2FFjYkiSizvZIqZ%2FE3VfJqY%3D)\r\n\r\n3. **文本嵌入（Embedding）层**：\r\n- 输入嵌入：`Token ID` → 稠密语义向量（索引查找）。\r\n- 输出嵌入：`隐藏状态` → 词表 `Logits`（矩阵乘法）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw18.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0RTDnZu0SO1CeAek7rPJ7s5he%2BA%3D)\r\n\r\n### 大模型模型结构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw19.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0oLFYwETFHgBhGA%2BtLUa3aQJzDA%3D)\r\n\r\n\r\n## 作业指引\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw20.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=43zzsnGBGlfotWicFAHotEu0%2B2w%3D)\r\n\r\n## 课后 Q&A\r\n\r\n**Q: 自注意力模块后的 `MLP` 和 `FFN` 是一个东西吗？**\r\n> A: 是的。`MLP`（多层感知器）与 `FFN`（前馈神经网络）在本课程语境下指代同一组件，即 Decoder Block 中的非线性变换层。\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**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\u002Fportal_news\u002F2026-xjq.png?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=oKlo6p8Ra0fcURk4F6PNTseTp%2BA%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=R03Psa4VwXKPuVbshfyqCCO33b4%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-63","study-63","2026 夏季 InfiniTensor 训练营通信与并行 1：《互联、集群与通信》",45,"\u002Fstudy-material\u002Fstudy-64","study-64","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-65","study-65","2026 夏季 InfiniTensor 训练营 Triton & 九齿 1：《Triton 练气术》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw01.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Z0gO%2FAENry9dvhWZT58a9cXI3oc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 16 日，2026 夏季 InfiniTensor 训练营推理引擎方向第二课《大模型原理与结构》开讲。\u003Cbr>推理引擎系列课程共 6 节，前置课程《AI 模型导论与推理基础原理》介绍了人工神经网络的基础知识，vec2vec 任务、矩阵乘、激活函数等内容，从本节课开始，正式进入大模型推理系统的核心领域。作为系列课程的第二讲，我们将深入拆解大模型的基石——Transformer 架构。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Ch2 id=\"为什么需要-transformer\">\u003Ca class=\"anchor\" href=\"#为什么需要-transformer\" aria-label=\"跳转到标题：为什么需要 Transformer？\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>为什么需要 Transformer？\u003C\u002Fh2>\u003Ch3 id=\"传统神经网络的局限\">\u003Ca class=\"anchor\" href=\"#传统神经网络的局限\" aria-label=\"跳转到标题：传统神经网络的局限\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>传统神经网络的局限\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>Vector to Vector\u003C\u002Fstrong>：输入输出维度固定，不便于直接处理变长序列（如手写数字识别）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>独立计算缺陷\u003C\u002Fstrong>：矩阵乘法中各输入向量相互独立，无法感知“道可道”中不同位置“道”字的语义差异。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw02.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TQ4aQNI4Q99qEY0Jck6l%2BrxIdRM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"sequence-to-sequence\">\u003Ca class=\"anchor\" href=\"#sequence-to-sequence\" aria-label=\"跳转到标题：Sequence to Sequence\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Sequence to Sequence\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>变长映射\u003C\u002Fstrong>：输入和输出均可为变长序列，二者长度不必相等（如翻译、文本生成、语音识别）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>上下文依赖\u003C\u002Fstrong>：词义随位置变化，需捕捉全局信息。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"早期解决方案及其瓶颈\">\u003Ca class=\"anchor\" href=\"#早期解决方案及其瓶颈\" aria-label=\"跳转到标题：早期解决方案及其瓶颈\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>早期解决方案及其瓶颈\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>滑动窗口\u003C\u002Fstrong>：仅能捕获局部固定范围上下文。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw03.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LMv%2BYEXo50E8el3Ote9TWOH8PUQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>\u003Cstrong>循环神经网络 (RNN)\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cul>\n\u003Cli>优点：理论支持无限长上下文（记忆模块 H 传递）。\u003C\u002Fli>\n\u003Cli>缺点：梯度消失\u002F信息衰减；串行计算导致硬件算力利用率低，无法并行。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw04.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FsnK85tDhxR%2BDM8D8NK29uuTvBI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"自注意力机制-self-attention\">\u003Ca class=\"anchor\" href=\"#自注意力机制-self-attention\" aria-label=\"跳转到标题：自注意力机制 (Self-Attention)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>自注意力机制 (Self-Attention)\u003C\u002Fh2>\u003Ch3 id=\"核心思想\">\u003Ca class=\"anchor\" href=\"#核心思想\" aria-label=\"跳转到标题：核心思想\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心思想\u003C\u002Fh3>\u003Cul>\n\u003Cli>上下文 = 当前输入与其他所有输入的相关性分布。\u003C\u002Fli>\n\u003Cli>突破窗口限制，实现全局信息交互。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw05.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=waAEgPSbUEDE3QgZsw9EP4W7ClM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"q-k-v-计算流程\">\u003Ca class=\"anchor\" href=\"#q-k-v-计算流程\" aria-label=\"跳转到标题：Q-K-V 计算流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>\u003Ccode>Q\u003C\u002Fcode>-\u003Ccode>K\u003C\u002Fcode>-\u003Ccode>V\u003C\u002Fcode> 计算流程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw06.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mZ3Xq%2FUYcBaJdvp%2Fv3nZHdU4DuI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>线性映射\u003C\u002Fstrong>：通过神经网络计算得到 \u003Ccode>Query\u003C\u002Fcode>、\u003Ccode>Key\u003C\u002Fcode>、\u003Ccode>Value\u003C\u002Fcode> 矩阵。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>注意力分数\u003C\u002Fstrong>：$alpha = q cdot k$\u003C\u002Fli>\n\u003Cli>\u003Cstrong>大模型算子 \u003Ccode>Softmax\u003C\u002Fcode>\u003C\u002Fstrong>：将分数转换为概率分布（和为 1）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>\u003Ccode>Safe Softmax\u003C\u002Fcode>\u003C\u002Fstrong>：减去最大值防止指数爆炸，保证数值稳定性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw07.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4yA2OtQ6f52qiFe6Y3U%2F3GkroyI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>因果掩码 (\u003Ccode>Causal Mask\u003C\u002Fcode>)\u003C\u002Fstrong>：主要用于自回归注意力，屏蔽未来 Token，防止训练时“作弊”。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw08.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sqbOmpclFeAC92TiFcxppW8u0K0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>矩阵并行化\u003C\u002Fstrong>：将逐对向量运算转化为矩阵乘法，充分利用 GPU 算力。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw09.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CfqDAZ9XPmGGctojxBHUkVrBIc8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"旋转位置编码-rope\">\u003Ca class=\"anchor\" href=\"#旋转位置编码-rope\" aria-label=\"跳转到标题：旋转位置编码 (RoPE)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>旋转位置编码 (\u003Ccode>RoPE\u003C\u002Fcode>)\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw10.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=q28hWd33PeYJ2ZEHsKAAtEzbKuE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>解决痛点\u003C\u002Fstrong>：不包含位置信息的自注意力具有排列等变性，本身无法区分 Token 的先后顺序，因此需要显式引入位置信息。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>原理\u003C\u002Fstrong>：通过绝对位置的三角函数（\u003Ccode>sin\u003C\u002Fcode>\u002F\u003Ccode>cos\u003C\u002Fcode>）对 \u003Ccode>Q\u003C\u002Fcode>、\u003Ccode>K\u003C\u002Fcode> 向量进行几何旋转。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>优势\u003C\u002Fstrong>：相对位置信息自然融入点积运算；角度表可预计算，推理高效。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"自注意力层\">\u003Ca class=\"anchor\" href=\"#自注意力层\" aria-label=\"跳转到标题：自注意力层\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>自注意力层\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw11.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QWuz5kIToJ8kF0N3wdJzI7fzCx0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"多头注意力\">\u003Ca class=\"anchor\" href=\"#多头注意力\" aria-label=\"跳转到标题：多头注意力\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>多头注意力\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw12.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fLOZhG0g8YSSxDL%2BNKgrSoCmIrk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>动机\u003C\u002Fstrong>：隐藏表示划分为多个注意力头，使模型能够在不同表示子空间中并行学习不同类型的语义关系和位置关系。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>机制\u003C\u002Fstrong>：多组独立的 \u003Ccode>Q\u003C\u002Fcode>、\u003Ccode>K\u003C\u002Fcode>、\u003Ccode>V\u003C\u002Fcode> 并行计算，捕获不同子空间的语义相关性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"解码器-decoder-模块\">\u003Ca class=\"anchor\" href=\"#解码器-decoder-模块\" aria-label=\"跳转到标题：解码器（Decoder）模块\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>解码器（Decoder）模块\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw13.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5oszdZoOeOzUYOVf3M9o2NQuu7E%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"decoder-模块标准组件\">\u003Ca class=\"anchor\" href=\"#decoder-模块标准组件\" aria-label=\"跳转到标题：Decoder 模块标准组件\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Decoder 模块标准组件\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>\u003Ccode>RMSNorm\u003C\u002Fcode>\u003C\u002Fstrong>：基于均方根的归一化方法，通过对隐藏状态进行尺度归一化，改善深层网络训练和推理过程中的数值稳定性。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>\u003Ccode>Residual\u003C\u002Fcode>\u003C\u002Fstrong>：残差连接，缓解梯度消失，保留原始信息。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>多层感知器 \u003Ccode>MLP\u003C\u002Fcode>\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw14.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HgJD9BYcIghd9%2BfsjsouN7W5UiQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"如何构建词表\">\u003Ca class=\"anchor\" href=\"#如何构建词表\" aria-label=\"跳转到标题：如何构建词表\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>如何构建词表\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>分词算法\u003C\u002Fstrong>：BPE（字节对编码）、WordPiece、Unigram；平衡词表大小与语义完整性。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw15.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=om42t9Lmvn%2F0snqHTI7VunU%2B5ro%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>\u003Cstrong>\u003Ccode>Tokenizer\u003C\u002Fcode>\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw16.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=H%2Bg1fbq2%2FDh4LwJLlZyidwzbHQQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw17.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=v3yOd%2FFjYkiSizvZIqZ%2FE3VfJqY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>\u003Cstrong>文本嵌入（Embedding）层\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cul>\n\u003Cli>输入嵌入：\u003Ccode>Token ID\u003C\u002Fcode> → 稠密语义向量（索引查找）。\u003C\u002Fli>\n\u003Cli>输出嵌入：\u003Ccode>隐藏状态\u003C\u002Fcode> → 词表 \u003Ccode>Logits\u003C\u002Fcode>（矩阵乘法）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw18.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0RTDnZu0SO1CeAek7rPJ7s5he%2BA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"大模型模型结构\">\u003Ca class=\"anchor\" href=\"#大模型模型结构\" aria-label=\"跳转到标题：大模型模型结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>大模型模型结构\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw19.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0oLFYwETFHgBhGA%2BtLUa3aQJzDA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"作业指引\">\u003Ca class=\"anchor\" href=\"#作业指引\" aria-label=\"跳转到标题：作业指引\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>作业指引\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_64_zw20.jpg?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=43zzsnGBGlfotWicFAHotEu0%2B2w%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"课后-q-amp-a\">\u003Ca class=\"anchor\" href=\"#课后-q-amp-a\" aria-label=\"跳转到标题：课后 Q&A\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>课后 Q&A\u003C\u002Fh2>\u003Cp>\u003Cstrong>Q: 自注意力模块后的 \u003Ccode>MLP\u003C\u002Fcode> 和 \u003Ccode>FFN\u003C\u002Fcode> 是一个东西吗？\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>A: 是的。\u003Ccode>MLP\u003C\u002Fcode>（多层感知器）与 \u003Ccode>FFN\u003C\u002Fcode>（前馈神经网络）在本课程语境下指代同一组件，即 Decoder Block 中的非线性变换层。\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>\u003Ch4 id=\"完整课程内容-请观看直播或查看课程回放\">\u003Ca class=\"anchor\" href=\"#完整课程内容-请观看直播或查看课程回放\" aria-label=\"跳转到标题：📺完整课程内容，请观看直播或查看课程回放：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📺完整课程内容，请观看直播或查看课程回放：\u003C\u002Fh4>\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\u002Fportal_news\u002F2026-xjq.png?Expires=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=oKlo6p8Ra0fcURk4F6PNTseTp%2BA%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=1792752557&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=R03Psa4VwXKPuVbshfyqCCO33b4%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,133,136,138,141,143,146,149,151,153,156,159,161,163,165,168,170,174],{"id":129,"text":130,"level":110},"为什么需要-transformer","为什么需要 Transformer？",{"id":132,"text":132,"level":9},"传统神经网络的局限",{"id":134,"text":135,"level":9},"sequence-to-sequence","Sequence to Sequence",{"id":137,"text":137,"level":9},"早期解决方案及其瓶颈",{"id":139,"text":140,"level":110},"自注意力机制-self-attention","自注意力机制 (Self-Attention)",{"id":142,"text":142,"level":9},"核心思想",{"id":144,"text":145,"level":9},"q-k-v-计算流程","Q-K-V 计算流程",{"id":147,"text":148,"level":9},"旋转位置编码-rope","旋转位置编码 (RoPE)",{"id":150,"text":150,"level":9},"自注意力层",{"id":152,"text":152,"level":9},"多头注意力",{"id":154,"text":155,"level":110},"解码器-decoder-模块","解码器（Decoder）模块",{"id":157,"text":158,"level":9},"decoder-模块标准组件","Decoder 模块标准组件",{"id":160,"text":160,"level":9},"如何构建词表",{"id":162,"text":162,"level":9},"大模型模型结构",{"id":164,"text":164,"level":110},"作业指引",{"id":166,"text":167,"level":110},"课后-q-amp-a","课后 Q&A",{"id":169,"text":169,"level":110},"加入训练营",{"id":171,"text":172,"level":173},"完整课程内容-请观看直播或查看课程回放","📺完整课程内容，请观看直播或查看课程回放：",4,{"id":175,"text":175,"level":110},"关注与交流"]