[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-46":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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ihPLIAuCbx%2FbInTQLbtnjbL%2FgEc%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RFJPol03COkEgvG5I%2FXurn1xrVs%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=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3EgCBk5kjXXxH6wOy5O8vDSNjzc%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_46_fm.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=zLXxAyH2%2BNUJLHjv2MjdyqVr%2Bpg%3D","InfiniTensor,冬季训练营,训练框架,编程基础,训练流程","讲解训练框架编程基础，覆盖训练流程拆解、模块组织与训练系统开发入门。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营训练系统第四课《训练框架编程基础》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw01.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=51NWVY%2Bgzk4AgyQ8f6rVa4Vd67M%3D)\r\n\r\n\r\n\r\n2026 年 1 月 29 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营训练系统系列第四课《训练框架编程基础》开讲。\r\n\r\n本节课侧重于深度学习训练相关的概念，以及实际使用 PyTorch 执行训练循环的整个流程。\r\n\r\n## PyTorch 概述\r\n\r\n### PyTorch简介：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw02.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=INAst4BgRDAfo4FjcgMTUMf0yaY%3D)\r\n\r\n### PyTorch 环境搭建与基础操作\r\n#### 1. 安装建议：\r\n- 推荐使用 Conda 或 venv 创建虚拟环境，避免版本冲突。\r\n- 国内用户安装包时切换清华源。\r\n- Google Colab 平台：提供免费 T4 GPU 算力，内置库安装完备。\r\n\r\n#### 2. Tensor 构造：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw03.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mcmiGOjdqF%2B3t9Xsxp3gPnmAVuw%3D)\r\n\r\n\r\n\r\n## Autograd 自动求导机制\r\n### 核心原理：\r\n1. 通过记录计算图实现自动求导，构建反向传播逻辑。\r\n2. 关键属性：`requires_grad=True` 的 Tensor 会记录所有计算操作。\r\n3. 中间节点（非叶子节点）：由叶子节点计算得到，保留反向求导算法于 `grad_fn` 属性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw04.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=B2rjBvv1x%2F0CY21Ehf7m3Hkd0uk%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw05.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1wvr9yD1Fod1%2FKYABCJuMxsH8Eo%3D)\r\n\r\n### 屏蔽 Autograd：\r\n1. 使用 `with torch.no_grad()` 上下文管理器屏蔽梯度追踪。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw06.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=M5Rd2VLyXHuzhZV096rH0ntE%2F90%3D)\r\n\r\n2. 使用 `.detach()` 方法获取不带梯度追踪的新 Tensor。\r\n\r\n## nn.Module 神经网络模块构建\r\n### 模块定义：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw07.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GnGuqg8AMpqFSyzBUu3PDuRX4Cw%3D)\r\n\r\n- 继承 `nn.Module` 类，声明子模块，定义 `forward` 方法\r\n- 快速实现顺序结构：使用 `nn.sequential`\r\n\r\n### 模型管理接口：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw08.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LlYaFjubCwQGZUdgwrCquHKBbGs%3D)\r\n\r\n- `parameters()`：返回需梯度更新的参数迭代器，用于传递给优化器。\r\n- `.to(device)`：将模型参数转移至指定设备。\r\n- 行为模式转换：`.train()` 与 `.eval()`，分别对应训练与推理模式。\r\n\r\n## nn.functional 无状态操作与损失函数\r\n### 无状态操作：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw09.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cwwhC8KnwZuULPNHaejSxss6iB8%3D)\r\n\r\n- 与 `nn.Module` 区别：无状态，不自动注册参数，需手动传入参数。\r\n- 典型应用：自定义前向传播逻辑，实现特殊网络结构。\r\n\r\n### 损失函数：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw10.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CgN29LYZOm3TMIw%2BNsee05d%2FKWU%3D)\r\n\r\n- 定义方式：通过 `nn.Module` 或 `nn.functional` 定义。\r\n- 常见损失函数：交叉熵、均方误差、负对数似然等。\r\n- 损失计算：通常输出为标量 Tensor。\r\n\r\n## 优化器与训练流程\r\n### 优化器（Optimizer）：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw11.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=o47zIusZkyqVKrhoP7%2Bs4r2HKbs%3D)\r\n\r\n- 作用：读取模型参数梯度，根据优化策略（如 SGD、Adam）更新参数。\r\n- 关键参数：学习率、动量、梯度衰减等。\r\n- 使用步骤：清零旧梯度（`zero_grad()`）、反向传播（`loss.backward()`）、参数更新（`optimizer.step()`）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw12.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VczJvYNLXhsVz%2FdJzb%2FtzDcgsTQ%3D)\r\n\r\n### 学习率调度器：\r\n- 作用：动态调整学习率，初期加速收敛，后期防止震荡。\r\n- 常见类型：StepLR、动态调整策略等。\r\n### 梯度裁剪与累积：\r\n- 梯度裁剪：限制梯度最大范数，增强训练稳定性。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw13.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DmPkgPRGJdh9xJAdEkYy5mSj6oo%3D)\r\n\r\n- 梯度累积：模拟大 batch 训练，缓解显存压力。\r\n\r\n## 完整训练循环示例\r\n### 训练准备：\r\n- 定义模型\r\n- 准备数据集与 DataLoader\r\n- 设置损失函数与优化器\r\n### 训练流程：\r\n- 模型设置为训练模式（`.train()`）。\r\n- 枚举每个 batch，执行前向传播、损失计算、反向传播、梯度裁剪（可选）、参数更新。\r\n- 验证阶段：模型设置为推理模式（`.eval()`），在`no_grad`上下文中计算验证损失与准确率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw14.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RRGrUkUufQiXXkW%2FRnd5JzSdQlg%3D)\r\n\r\n## 高级主题与扩展\r\n### 模型 checkpointing：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw15.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tKojmDiqdH3QlguFupKAsMx72yI%3D)\r\n\r\n\r\n\r\n### 混合精度训练：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw16.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=By1f67CLlDDAsaYI2Im%2B4PjcGVs%3D)\r\n\r\n- 使用 `Autocast` 与 `GradScaler` 提升计算效率与减少显存占用。\r\n### Torchvision 与 Huggingface 应用：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw17.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b1KlJFqVRtdaaD1KQAlnlXXuOsc%3D)\r\n\r\n## Megatron-LM 框架介绍：\r\n- 背景：应对大模型训练挑战，如显存限制、计算瓶颈等。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw18.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Bbtp4iiclvsmVelA3wIftWgMk0A%3D)\r\n\r\n- 核心特性：多种并行策略（张量并行、流水线并行等）、混合精度策略、分布式特性等。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw19.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HTZaLXiGp78hAG2vsF0PazqE4Lc%3D)\r\n\r\n\r\n- 使用流程：基于 NGC Docker 环境；使用官方示例脚本（如 BERT、GPT）进行预训练；关键训练参数配置。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw20.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bu0q2srjKq4GqLn3rOombJeMhmE%3D)\r\n\r\n\r\n\r\n## 总结\r\n本节课通过理论讲解与代码示例相结合，掌握从零开始构建和训练 PyTorch 模型的完整技能栈，并为应对大尺度模型训练的挑战奠定了基础。\r\n\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_kcb.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6s0mlm18SaIAbSmWzBUsWpRy3BA%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_ewm.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=73i%2BTxWax0UR1WFUTumK55SpDGo%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-45","study-45","InfiniTensor 冬季训练营前沿模型第五课《大模型自主智能体》",63,"\u002Fstudy-material\u002Fstudy-46","study-46","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-47","study-47","InfiniTensor 冬季训练营推理引擎第五课《大模型服务系统》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw01.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=51NWVY%2Bgzk4AgyQ8f6rVa4Vd67M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 29 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营训练系统系列第四课《训练框架编程基础》开讲。\u003C\u002Fp>\n\u003Cp>本节课侧重于深度学习训练相关的概念，以及实际使用 PyTorch 执行训练循环的整个流程。\u003C\u002Fp>\n\u003Ch2 id=\"pytorch-概述\">\u003Ca class=\"anchor\" href=\"#pytorch-概述\" aria-label=\"跳转到标题：PyTorch 概述\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PyTorch 概述\u003C\u002Fh2>\u003Ch3 id=\"pytorch简介\">\u003Ca class=\"anchor\" href=\"#pytorch简介\" aria-label=\"跳转到标题：PyTorch简介：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PyTorch简介：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw02.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=INAst4BgRDAfo4FjcgMTUMf0yaY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"pytorch-环境搭建与基础操作\">\u003Ca class=\"anchor\" href=\"#pytorch-环境搭建与基础操作\" aria-label=\"跳转到标题：PyTorch 环境搭建与基础操作\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PyTorch 环境搭建与基础操作\u003C\u002Fh3>\u003Ch4 id=\"1-安装建议\">\u003Ca class=\"anchor\" href=\"#1-安装建议\" aria-label=\"跳转到标题：1. 安装建议：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 安装建议：\u003C\u002Fh4>\u003Cul>\n\u003Cli>推荐使用 Conda 或 venv 创建虚拟环境，避免版本冲突。\u003C\u002Fli>\n\u003Cli>国内用户安装包时切换清华源。\u003C\u002Fli>\n\u003Cli>Google Colab 平台：提供免费 T4 GPU 算力，内置库安装完备。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-tensor-构造\">\u003Ca class=\"anchor\" href=\"#2-tensor-构造\" aria-label=\"跳转到标题：2. Tensor 构造：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. Tensor 构造：\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw03.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mcmiGOjdqF%2B3t9Xsxp3gPnmAVuw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"autograd-自动求导机制\">\u003Ca class=\"anchor\" href=\"#autograd-自动求导机制\" aria-label=\"跳转到标题：Autograd 自动求导机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Autograd 自动求导机制\u003C\u002Fh2>\u003Ch3 id=\"核心原理\">\u003Ca class=\"anchor\" href=\"#核心原理\" aria-label=\"跳转到标题：核心原理：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心原理：\u003C\u002Fh3>\u003Col>\n\u003Cli>通过记录计算图实现自动求导，构建反向传播逻辑。\u003C\u002Fli>\n\u003Cli>关键属性：\u003Ccode>requires_grad=True\u003C\u002Fcode> 的 Tensor 会记录所有计算操作。\u003C\u002Fli>\n\u003Cli>中间节点（非叶子节点）：由叶子节点计算得到，保留反向求导算法于 \u003Ccode>grad_fn\u003C\u002Fcode> 属性。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw04.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=B2rjBvv1x%2F0CY21Ehf7m3Hkd0uk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw05.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1wvr9yD1Fod1%2FKYABCJuMxsH8Eo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"屏蔽-autograd\">\u003Ca class=\"anchor\" href=\"#屏蔽-autograd\" aria-label=\"跳转到标题：屏蔽 Autograd：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>屏蔽 Autograd：\u003C\u002Fh3>\u003Col>\n\u003Cli>使用 \u003Ccode>with torch.no_grad()\u003C\u002Fcode> 上下文管理器屏蔽梯度追踪。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw06.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=M5Rd2VLyXHuzhZV096rH0ntE%2F90%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>使用 \u003Ccode>.detach()\u003C\u002Fcode> 方法获取不带梯度追踪的新 Tensor。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2 id=\"nn-module-神经网络模块构建\">\u003Ca class=\"anchor\" href=\"#nn-module-神经网络模块构建\" aria-label=\"跳转到标题：nn.Module 神经网络模块构建\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>nn.Module 神经网络模块构建\u003C\u002Fh2>\u003Ch3 id=\"模块定义\">\u003Ca class=\"anchor\" href=\"#模块定义\" aria-label=\"跳转到标题：模块定义：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模块定义：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw07.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GnGuqg8AMpqFSyzBUu3PDuRX4Cw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>继承 \u003Ccode>nn.Module\u003C\u002Fcode> 类，声明子模块，定义 \u003Ccode>forward\u003C\u002Fcode> 方法\u003C\u002Fli>\n\u003Cli>快速实现顺序结构：使用 \u003Ccode>nn.sequential\u003C\u002Fcode>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"模型管理接口\">\u003Ca class=\"anchor\" href=\"#模型管理接口\" aria-label=\"跳转到标题：模型管理接口：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型管理接口：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw08.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LlYaFjubCwQGZUdgwrCquHKBbGs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Ccode>parameters()\u003C\u002Fcode>：返回需梯度更新的参数迭代器，用于传递给优化器。\u003C\u002Fli>\n\u003Cli>\u003Ccode>.to(device)\u003C\u002Fcode>：将模型参数转移至指定设备。\u003C\u002Fli>\n\u003Cli>行为模式转换：\u003Ccode>.train()\u003C\u002Fcode> 与 \u003Ccode>.eval()\u003C\u002Fcode>，分别对应训练与推理模式。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"nn-functional-无状态操作与损失函数\">\u003Ca class=\"anchor\" href=\"#nn-functional-无状态操作与损失函数\" aria-label=\"跳转到标题：nn.functional 无状态操作与损失函数\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>nn.functional 无状态操作与损失函数\u003C\u002Fh2>\u003Ch3 id=\"无状态操作\">\u003Ca class=\"anchor\" href=\"#无状态操作\" aria-label=\"跳转到标题：无状态操作：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>无状态操作：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw09.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cwwhC8KnwZuULPNHaejSxss6iB8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>与 \u003Ccode>nn.Module\u003C\u002Fcode> 区别：无状态，不自动注册参数，需手动传入参数。\u003C\u002Fli>\n\u003Cli>典型应用：自定义前向传播逻辑，实现特殊网络结构。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"损失函数\">\u003Ca class=\"anchor\" href=\"#损失函数\" aria-label=\"跳转到标题：损失函数：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>损失函数：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw10.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CgN29LYZOm3TMIw%2BNsee05d%2FKWU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>定义方式：通过 \u003Ccode>nn.Module\u003C\u002Fcode> 或 \u003Ccode>nn.functional\u003C\u002Fcode> 定义。\u003C\u002Fli>\n\u003Cli>常见损失函数：交叉熵、均方误差、负对数似然等。\u003C\u002Fli>\n\u003Cli>损失计算：通常输出为标量 Tensor。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"优化器与训练流程\">\u003Ca class=\"anchor\" href=\"#优化器与训练流程\" aria-label=\"跳转到标题：优化器与训练流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化器与训练流程\u003C\u002Fh2>\u003Ch3 id=\"优化器-optimizer\">\u003Ca class=\"anchor\" href=\"#优化器-optimizer\" aria-label=\"跳转到标题：优化器（Optimizer）：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化器（Optimizer）：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw11.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=o47zIusZkyqVKrhoP7%2Bs4r2HKbs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>作用：读取模型参数梯度，根据优化策略（如 SGD、Adam）更新参数。\u003C\u002Fli>\n\u003Cli>关键参数：学习率、动量、梯度衰减等。\u003C\u002Fli>\n\u003Cli>使用步骤：清零旧梯度（\u003Ccode>zero_grad()\u003C\u002Fcode>）、反向传播（\u003Ccode>loss.backward()\u003C\u002Fcode>）、参数更新（\u003Ccode>optimizer.step()\u003C\u002Fcode>）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw12.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VczJvYNLXhsVz%2FdJzb%2FtzDcgsTQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"学习率调度器\">\u003Ca class=\"anchor\" href=\"#学习率调度器\" aria-label=\"跳转到标题：学习率调度器：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>学习率调度器：\u003C\u002Fh3>\u003Cul>\n\u003Cli>作用：动态调整学习率，初期加速收敛，后期防止震荡。\u003C\u002Fli>\n\u003Cli>常见类型：StepLR、动态调整策略等。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"梯度裁剪与累积\">\u003Ca class=\"anchor\" href=\"#梯度裁剪与累积\" aria-label=\"跳转到标题：梯度裁剪与累积：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>梯度裁剪与累积：\u003C\u002Fh3>\u003Cul>\n\u003Cli>梯度裁剪：限制梯度最大范数，增强训练稳定性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw13.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DmPkgPRGJdh9xJAdEkYy5mSj6oo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>梯度累积：模拟大 batch 训练，缓解显存压力。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"完整训练循环示例\">\u003Ca class=\"anchor\" href=\"#完整训练循环示例\" aria-label=\"跳转到标题：完整训练循环示例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>完整训练循环示例\u003C\u002Fh2>\u003Ch3 id=\"训练准备\">\u003Ca class=\"anchor\" href=\"#训练准备\" aria-label=\"跳转到标题：训练准备：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>训练准备：\u003C\u002Fh3>\u003Cul>\n\u003Cli>定义模型\u003C\u002Fli>\n\u003Cli>准备数据集与 DataLoader\u003C\u002Fli>\n\u003Cli>设置损失函数与优化器\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"训练流程\">\u003Ca class=\"anchor\" href=\"#训练流程\" aria-label=\"跳转到标题：训练流程：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>训练流程：\u003C\u002Fh3>\u003Cul>\n\u003Cli>模型设置为训练模式（\u003Ccode>.train()\u003C\u002Fcode>）。\u003C\u002Fli>\n\u003Cli>枚举每个 batch，执行前向传播、损失计算、反向传播、梯度裁剪（可选）、参数更新。\u003C\u002Fli>\n\u003Cli>验证阶段：模型设置为推理模式（\u003Ccode>.eval()\u003C\u002Fcode>），在\u003Ccode>no_grad\u003C\u002Fcode>上下文中计算验证损失与准确率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw14.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RRGrUkUufQiXXkW%2FRnd5JzSdQlg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"高级主题与扩展\">\u003Ca class=\"anchor\" href=\"#高级主题与扩展\" aria-label=\"跳转到标题：高级主题与扩展\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>高级主题与扩展\u003C\u002Fh2>\u003Ch3 id=\"模型-checkpointing\">\u003Ca class=\"anchor\" href=\"#模型-checkpointing\" aria-label=\"跳转到标题：模型 checkpointing：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型 checkpointing：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw15.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tKojmDiqdH3QlguFupKAsMx72yI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"混合精度训练\">\u003Ca class=\"anchor\" href=\"#混合精度训练\" aria-label=\"跳转到标题：混合精度训练：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>混合精度训练：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw16.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=By1f67CLlDDAsaYI2Im%2B4PjcGVs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>使用 \u003Ccode>Autocast\u003C\u002Fcode> 与 \u003Ccode>GradScaler\u003C\u002Fcode> 提升计算效率与减少显存占用。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"torchvision-与-huggingface-应用\">\u003Ca class=\"anchor\" href=\"#torchvision-与-huggingface-应用\" aria-label=\"跳转到标题：Torchvision 与 Huggingface 应用：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Torchvision 与 Huggingface 应用：\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw17.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b1KlJFqVRtdaaD1KQAlnlXXuOsc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"megatron-lm-框架介绍\">\u003Ca class=\"anchor\" href=\"#megatron-lm-框架介绍\" aria-label=\"跳转到标题：Megatron-LM 框架介绍：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Megatron-LM 框架介绍：\u003C\u002Fh2>\u003Cul>\n\u003Cli>背景：应对大模型训练挑战，如显存限制、计算瓶颈等。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw18.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Bbtp4iiclvsmVelA3wIftWgMk0A%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>核心特性：多种并行策略（张量并行、流水线并行等）、混合精度策略、分布式特性等。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw19.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HTZaLXiGp78hAG2vsF0PazqE4Lc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>使用流程：基于 NGC Docker 环境；使用官方示例脚本（如 BERT、GPT）进行预训练；关键训练参数配置。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_46_zw20.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bu0q2srjKq4GqLn3rOombJeMhmE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 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\u002Fh2>\u003Cp>本节课通过理论讲解与代码示例相结合，掌握从零开始构建和训练 PyTorch 模型的完整技能栈，并为应对大尺度模型训练的挑战奠定了基础。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6s0mlm18SaIAbSmWzBUsWpRy3BA%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_ewm.jpg?Expires=1792752556&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=73i%2BTxWax0UR1WFUTumK55SpDGo%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,134,137,141,144,147,150,153,156,159,162,165,168,171,173,176,179,182,184,187,190,192,195,198,201,204,206,209],{"id":129,"text":130,"level":110},"pytorch-概述","PyTorch 概述",{"id":132,"text":133,"level":9},"pytorch简介","PyTorch简介：",{"id":135,"text":136,"level":9},"pytorch-环境搭建与基础操作","PyTorch 环境搭建与基础操作",{"id":138,"text":139,"level":140},"1-安装建议","1. 安装建议：",4,{"id":142,"text":143,"level":140},"2-tensor-构造","2. Tensor 构造：",{"id":145,"text":146,"level":110},"autograd-自动求导机制","Autograd 自动求导机制",{"id":148,"text":149,"level":9},"核心原理","核心原理：",{"id":151,"text":152,"level":9},"屏蔽-autograd","屏蔽 Autograd：",{"id":154,"text":155,"level":110},"nn-module-神经网络模块构建","nn.Module 神经网络模块构建",{"id":157,"text":158,"level":9},"模块定义","模块定义：",{"id":160,"text":161,"level":9},"模型管理接口","模型管理接口：",{"id":163,"text":164,"level":110},"nn-functional-无状态操作与损失函数","nn.functional 无状态操作与损失函数",{"id":166,"text":167,"level":9},"无状态操作","无状态操作：",{"id":169,"text":170,"level":9},"损失函数","损失函数：",{"id":172,"text":172,"level":110},"优化器与训练流程",{"id":174,"text":175,"level":9},"优化器-optimizer","优化器（Optimizer）：",{"id":177,"text":178,"level":9},"学习率调度器","学习率调度器：",{"id":180,"text":181,"level":9},"梯度裁剪与累积","梯度裁剪与累积：",{"id":183,"text":183,"level":110},"完整训练循环示例",{"id":185,"text":186,"level":9},"训练准备","训练准备：",{"id":188,"text":189,"level":9},"训练流程","训练流程：",{"id":191,"text":191,"level":110},"高级主题与扩展",{"id":193,"text":194,"level":9},"模型-checkpointing","模型 checkpointing：",{"id":196,"text":197,"level":9},"混合精度训练","混合精度训练：",{"id":199,"text":200,"level":9},"torchvision-与-huggingface-应用","Torchvision 与 Huggingface 应用：",{"id":202,"text":203,"level":110},"megatron-lm-框架介绍","Megatron-LM 框架介绍：",{"id":205,"text":205,"level":110},"总结",{"id":207,"text":208,"level":9},"学习贴士","💡 学习贴士",{"id":210,"text":211,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]