[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-79":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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=nCcM9ECgK5PlmraUwj%2BH%2F1S%2BELk%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kRi1yuIH5NG2docsx57MXWlzm1k%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eT9ZAOmPWlOldw%2FMHXc0X0x2%2Bhg%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_79_fm.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DfuUtpfjctDzD0pVTd3r8qCvRnI%3D","KV Cache,大模型推理,对话系统,采样策略,Chat Template","讲解自回归文本生成，核心拆解 KV Cache Prefill\u002FDecode 双阶段加速，介绍 Chat Template、角色对话结构，对比 Top-P\u002FBeam Search 等各类采样策略。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营推理引擎 3：《从大模型推理到 AI 对话》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1FJe0Cc8%2BmQNTlcSjLEgWMAEQ8A%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 24 日，2026 夏季 InfiniTensor 训练营推理引擎方向第三课《从大模型推理到 AI 对话》开讲。\r\n前两课依次讲解神经网络基础、Seq2Seq 任务演进，完整拆解 Transformer 自注意力、RoPE、多头注意力、Decoder 模块、分词与嵌入层底层结构，为本节推理与对话工程落地打好理论基础。\u003C\u002Fspan>\r\n\r\n本节课承接前两课 Transformer 底层结构知识，讲解因果语言模型生成逻辑、KV Cache 核心加速机制，以及 Chat Template、采样策略、模型文件等对话推理完整落地知识。\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hpclxZsmrFXpbuS8LeD7U5he430%3D)\r\n\r\n## 因果语言模型：文本生成\r\n### 分类任务到文本生成转换\r\n词性标注属于序列标注任务，每个输入 Token 对应一个类别标签；而故事续写、问答对话属于自回归文本生成任务。\r\n\r\n实现思路：将下一个 Token 的预测转为词表范围内分类，词表有限则生成可能性可控，每次仅预测单个 Token，循环迭代输出完整文本。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rOrDv%2F%2FSi29vblNmjTUfU1DxOsg%3D)\r\n\r\n### 逐 Token 多轮生成流程\r\n1. 输入一段上下文文本送入模型；\r\n2. 模型计算全部上下文特征，输出词表维度概率分布；\r\n3. 采样选出概率最高\u002F符合随机策略的下一个 Token；\r\n4. 将新 Token 拼接至原始上下文，作为下一轮输入；\r\n5. 循环直至模型输出结束标记 `\u003C\u002Fs>` 终止生成。\r\n\r\n### 原生推理性能缺陷\r\n无缓存机制时，每轮生成都需要重新计算历史 Token 的 K\u002FV，并重新计算完整 Attention，随着上下文长度增加，单步计算成本持续增长。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ff2mJWJmj2Gs2%2FsA0zcTNi5%2BPIg%3D)\r\n\r\n## KV Cache 推理核心加速机制\r\n### 核心原理\r\n每一轮推理保存所有历史 Token 的 K \u002F V 向量缓存，后续生成新 Token 时仅计算当前 QKV，复用历史缓存，计算复杂度降至 O(N)。\r\n- 历史 K、V 无需重复矩阵乘；\r\n- 仅新增当前 Token 的 Q、K、V，并将新的 K\u002FV 写入缓存；计算 Attention 时使用当前 Q 与历史缓存的全部 K\u002FV 进行交互。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rUM%2BSpjo8jbBf0BYjhLTv1Ex4wM%3D)\r\n\r\n### 两大推理阶段划分\r\n#### 1.Prefill 阶段\r\nPrefill 阶段一次性处理完整输入上下文，计算并缓存所有历史 Token 的 KV，同时生成首个待生成 Token 的概率分布，为后续 Decode 阶段提供起点。\r\n#### 2.Decode 阶段\r\nDecode 阶段每轮输入当前待计算的单个 Token，同时读取历史 KV Cache，仅计算该 Token 对历史上下文的注意力结果，并生成下一个 Token。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bVJ4zsCrwRjQdbbG61DbkYKMK4c%3D)\r\n\r\n## AI 对话系统实现方案\r\n### 对话本质：带角色标识的文本续写\r\n普通文本续写无身份区分，对话需要区分 User \u002F Assistant 两类发言主体，通过专用 Tag 标记角色边界，让模型识别交互逻辑。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yxalsS%2B79eVR%2B%2Fa6%2FsSpwXDNSks%3D)\r\n\r\n### Messages 对话数据结构\r\n统一对话存储格式，每条消息包含两个核心字段：\r\n- `role`：角色（user \u002F assistant \u002F system）；\r\n- `content`：对应文本内容。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D1%2BWKmY1QVDaA6n1mnBZRWdCkGU%3D)\r\n\r\n### Chat Template 对话模板引擎\r\n基于模板语言渲染 Messages 列表为模型可识别完整文本，内置循环、判断逻辑，自动拼接各类角色 Tag。\r\n1. System 系统提示词：置于文本最前端，约束模型回答风格、安全边界、输出格式；\r\n2. 思考专用 Tag：部分模型支持前置思考段，可手动截断强制结束思考；\r\n3. Tag 需匹配模型训练阶段使用的模板格式，否则模型可能无法正确理解角色边界，导致生成质量下降或输出异常。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J9hJ8wWNWmqGOwjyB7m%2BxxFyzVU%3D)\r\n\r\n## 生成采样策略详解\r\n模型输出原始结果为 Logits 原始分值，经 Softmax 转为概率分布，通过采样算法选取下一个 Token。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TsTTwWtgyCMge1lrSW37hWLSK5Q%3D)\r\n\r\n### Argmax 采样\r\n永久选取概率最大值 Token，输出固定、无多样性，易出现循环重复语句。\r\n\r\n### 随机基础采样\r\n- 基于概率加权随机选取，相同输入可产出不同回答；\r\n- 但低概率非法词汇有被选中风险。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LAnIDJ3ff6uyo%2BVbXne9KYIeUJc%3D)\r\n\r\n### Top-K 采样\r\n- 仅保留概率排名前 K 的 Token 作为候选池，过滤极低概率无效词汇；\r\n- 缺陷是固定 K 值无法适配各类文本分布。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ITTpk1UjX0eAlSRjU1KGE9ShT1U%3D)\r\n\r\n### Top-P 核采样\r\n累加概率至阈值 P，动态截取候选集合，灵活适配稀疏\u002F密集概率分布，工程常用方案。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw13.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pjgG4FCrdbB4wAlsOKB55F2gSg0%3D)\r\n\r\n### Temperature 温度系数\r\n缩放 Logits 分布：\r\n- T > 1：分布平滑，多样性提升；\r\n- T → 0：分布陡峭，趋近 Argmax。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw14.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=t8REnNAPdcgmlrrq7kKGxwIL2FI%3D)\r\n\r\n### Beam Search\r\n- 保留多条高概率生成分支全局择优，文本流畅度更高；\r\n- 缺点是内存、算力开销大，线上推理极少使用。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw15.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xpIhMTelP0%2F0WIGWUvO8qiW%2BshI%3D)\r\n\r\n## Hugging Face 与 Transformers 库\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw16.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SrHJsrDjJoXSg8g8SKv6e6HWQR4%3D)\r\n\r\n1. `config.json`：存储模型全局超参数（隐藏层维度、头数、MLP 尺寸等）；\r\n2. Safetensors 以安全、高效的方式保存模型张量数据，通过参数名称映射到模型结构中的对应权重。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw17.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y4OYR5zR4QmRO9TxY77o4X%2Bu1yE%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw18.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MedvDk%2B5u7LV7hAsWOGhpBu%2FuS0%3D)\r\n\r\n\r\n\r\n## 课程作业提示\r\n1. 仔细阅读作业 README，所有实现步骤、约束均有完整说明；\r\n2. 作业难度较高，建议尽早动手实践，疑问可在训练营官方社群咨询；\r\n3. 后续课程会安排专项作业答疑与完整项目讲解。\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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j5LLQWJGl7lfBWs0J77BjrIj89Q%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5J3BUGj%2FbNiiRCskyxeUoMVLor4%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-78","study-78","2026 夏季 InfiniTensor 训练营 Triton & 九齿 3：《九齿一重：初窥》",30,"\u002Fstudy-material\u002Fstudy-79","study-79","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-80","study-80","2026 夏季 InfiniTensor 训练营 CUDA 编程 4：《CUDA 基础入门：向量规约》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1FJe0Cc8%2BmQNTlcSjLEgWMAEQ8A%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 24 日，2026 夏季 InfiniTensor 训练营推理引擎方向第三课《从大模型推理到 AI 对话》开讲。\u003Cbr>前两课依次讲解神经网络基础、Seq2Seq 任务演进，完整拆解 Transformer 自注意力、RoPE、多头注意力、Decoder 模块、分词与嵌入层底层结构，为本节推理与对话工程落地打好理论基础。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课承接前两课 Transformer 底层结构知识，讲解因果语言模型生成逻辑、KV Cache 核心加速机制，以及 Chat Template、采样策略、模型文件等对话推理完整落地知识。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hpclxZsmrFXpbuS8LeD7U5he430%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"因果语言模型-文本生成\">\u003Ca class=\"anchor\" href=\"#因果语言模型-文本生成\" aria-label=\"跳转到标题：因果语言模型：文本生成\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>因果语言模型：文本生成\u003C\u002Fh2>\u003Ch3 id=\"分类任务到文本生成转换\">\u003Ca class=\"anchor\" href=\"#分类任务到文本生成转换\" aria-label=\"跳转到标题：分类任务到文本生成转换\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>分类任务到文本生成转换\u003C\u002Fh3>\u003Cp>词性标注属于序列标注任务，每个输入 Token 对应一个类别标签；而故事续写、问答对话属于自回归文本生成任务。\u003C\u002Fp>\n\u003Cp>实现思路：将下一个 Token 的预测转为词表范围内分类，词表有限则生成可能性可控，每次仅预测单个 Token，循环迭代输出完整文本。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rOrDv%2F%2FSi29vblNmjTUfU1DxOsg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"逐-token-多轮生成流程\">\u003Ca class=\"anchor\" href=\"#逐-token-多轮生成流程\" aria-label=\"跳转到标题：逐 Token 多轮生成流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>逐 Token 多轮生成流程\u003C\u002Fh3>\u003Col>\n\u003Cli>输入一段上下文文本送入模型；\u003C\u002Fli>\n\u003Cli>模型计算全部上下文特征，输出词表维度概率分布；\u003C\u002Fli>\n\u003Cli>采样选出概率最高\u002F符合随机策略的下一个 Token；\u003C\u002Fli>\n\u003Cli>将新 Token 拼接至原始上下文，作为下一轮输入；\u003C\u002Fli>\n\u003Cli>循环直至模型输出结束标记 \u003Ccode><\u002Fs>\u003C\u002Fcode> 终止生成。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"原生推理性能缺陷\">\u003Ca class=\"anchor\" href=\"#原生推理性能缺陷\" aria-label=\"跳转到标题：原生推理性能缺陷\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>原生推理性能缺陷\u003C\u002Fh3>\u003Cp>无缓存机制时，每轮生成都需要重新计算历史 Token 的 K\u002FV，并重新计算完整 Attention，随着上下文长度增加，单步计算成本持续增长。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ff2mJWJmj2Gs2%2FsA0zcTNi5%2BPIg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"kv-cache-推理核心加速机制\">\u003Ca class=\"anchor\" href=\"#kv-cache-推理核心加速机制\" aria-label=\"跳转到标题：KV Cache 推理核心加速机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>KV Cache 推理核心加速机制\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>\u003Cp>每一轮推理保存所有历史 Token 的 K \u002F V 向量缓存，后续生成新 Token 时仅计算当前 QKV，复用历史缓存，计算复杂度降至 O(N)。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>历史 K、V 无需重复矩阵乘；\u003C\u002Fli>\n\u003Cli>仅新增当前 Token 的 Q、K、V，并将新的 K\u002FV 写入缓存；计算 Attention 时使用当前 Q 与历史缓存的全部 K\u002FV 进行交互。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rUM%2BSpjo8jbBf0BYjhLTv1Ex4wM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"两大推理阶段划分\">\u003Ca class=\"anchor\" href=\"#两大推理阶段划分\" aria-label=\"跳转到标题：两大推理阶段划分\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>两大推理阶段划分\u003C\u002Fh3>\u003Ch4 id=\"1-prefill-阶段\">\u003Ca class=\"anchor\" href=\"#1-prefill-阶段\" aria-label=\"跳转到标题：1.Prefill 阶段\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.Prefill 阶段\u003C\u002Fh4>\u003Cp>Prefill 阶段一次性处理完整输入上下文，计算并缓存所有历史 Token 的 KV，同时生成首个待生成 Token 的概率分布，为后续 Decode 阶段提供起点。\u003C\u002Fp>\n\u003Ch4 id=\"2-decode-阶段\">\u003Ca class=\"anchor\" href=\"#2-decode-阶段\" aria-label=\"跳转到标题：2.Decode 阶段\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.Decode 阶段\u003C\u002Fh4>\u003Cp>Decode 阶段每轮输入当前待计算的单个 Token，同时读取历史 KV Cache，仅计算该 Token 对历史上下文的注意力结果，并生成下一个 Token。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bVJ4zsCrwRjQdbbG61DbkYKMK4c%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"ai-对话系统实现方案\">\u003Ca class=\"anchor\" href=\"#ai-对话系统实现方案\" aria-label=\"跳转到标题：AI 对话系统实现方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>AI 对话系统实现方案\u003C\u002Fh2>\u003Ch3 id=\"对话本质-带角色标识的文本续写\">\u003Ca class=\"anchor\" href=\"#对话本质-带角色标识的文本续写\" aria-label=\"跳转到标题：对话本质：带角色标识的文本续写\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>对话本质：带角色标识的文本续写\u003C\u002Fh3>\u003Cp>普通文本续写无身份区分，对话需要区分 User \u002F Assistant 两类发言主体，通过专用 Tag 标记角色边界，让模型识别交互逻辑。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yxalsS%2B79eVR%2B%2Fa6%2FsSpwXDNSks%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"messages-对话数据结构\">\u003Ca class=\"anchor\" href=\"#messages-对话数据结构\" aria-label=\"跳转到标题：Messages 对话数据结构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Messages 对话数据结构\u003C\u002Fh3>\u003Cp>统一对话存储格式，每条消息包含两个核心字段：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Ccode>role\u003C\u002Fcode>：角色（user \u002F assistant \u002F system）；\u003C\u002Fli>\n\u003Cli>\u003Ccode>content\u003C\u002Fcode>：对应文本内容。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D1%2BWKmY1QVDaA6n1mnBZRWdCkGU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"chat-template-对话模板引擎\">\u003Ca class=\"anchor\" href=\"#chat-template-对话模板引擎\" aria-label=\"跳转到标题：Chat Template 对话模板引擎\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Chat Template 对话模板引擎\u003C\u002Fh3>\u003Cp>基于模板语言渲染 Messages 列表为模型可识别完整文本，内置循环、判断逻辑，自动拼接各类角色 Tag。\u003C\u002Fp>\n\u003Col>\n\u003Cli>System 系统提示词：置于文本最前端，约束模型回答风格、安全边界、输出格式；\u003C\u002Fli>\n\u003Cli>思考专用 Tag：部分模型支持前置思考段，可手动截断强制结束思考；\u003C\u002Fli>\n\u003Cli>Tag 需匹配模型训练阶段使用的模板格式，否则模型可能无法正确理解角色边界，导致生成质量下降或输出异常。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J9hJ8wWNWmqGOwjyB7m%2BxxFyzVU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"生成采样策略详解\">\u003Ca class=\"anchor\" href=\"#生成采样策略详解\" aria-label=\"跳转到标题：生成采样策略详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>生成采样策略详解\u003C\u002Fh2>\u003Cp>模型输出原始结果为 Logits 原始分值，经 Softmax 转为概率分布，通过采样算法选取下一个 Token。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TsTTwWtgyCMge1lrSW37hWLSK5Q%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"argmax-采样\">\u003Ca class=\"anchor\" href=\"#argmax-采样\" aria-label=\"跳转到标题：Argmax 采样\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Argmax 采样\u003C\u002Fh3>\u003Cp>永久选取概率最大值 Token，输出固定、无多样性，易出现循环重复语句。\u003C\u002Fp>\n\u003Ch3 id=\"随机基础采样\">\u003Ca class=\"anchor\" href=\"#随机基础采样\" aria-label=\"跳转到标题：随机基础采样\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>随机基础采样\u003C\u002Fh3>\u003Cul>\n\u003Cli>基于概率加权随机选取，相同输入可产出不同回答；\u003C\u002Fli>\n\u003Cli>但低概率非法词汇有被选中风险。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LAnIDJ3ff6uyo%2BVbXne9KYIeUJc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"top-k-采样\">\u003Ca class=\"anchor\" href=\"#top-k-采样\" aria-label=\"跳转到标题：Top-K 采样\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Top-K 采样\u003C\u002Fh3>\u003Cul>\n\u003Cli>仅保留概率排名前 K 的 Token 作为候选池，过滤极低概率无效词汇；\u003C\u002Fli>\n\u003Cli>缺陷是固定 K 值无法适配各类文本分布。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ITTpk1UjX0eAlSRjU1KGE9ShT1U%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"top-p-核采样\">\u003Ca class=\"anchor\" href=\"#top-p-核采样\" aria-label=\"跳转到标题：Top-P 核采样\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Top-P 核采样\u003C\u002Fh3>\u003Cp>累加概率至阈值 P，动态截取候选集合，灵活适配稀疏\u002F密集概率分布，工程常用方案。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw13.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pjgG4FCrdbB4wAlsOKB55F2gSg0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"temperature-温度系数\">\u003Ca class=\"anchor\" href=\"#temperature-温度系数\" aria-label=\"跳转到标题：Temperature 温度系数\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Temperature 温度系数\u003C\u002Fh3>\u003Cp>缩放 Logits 分布：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>T > 1：分布平滑，多样性提升；\u003C\u002Fli>\n\u003Cli>T → 0：分布陡峭，趋近 Argmax。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw14.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=t8REnNAPdcgmlrrq7kKGxwIL2FI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"beam-search\">\u003Ca class=\"anchor\" href=\"#beam-search\" aria-label=\"跳转到标题：Beam Search\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Beam Search\u003C\u002Fh3>\u003Cul>\n\u003Cli>保留多条高概率生成分支全局择优，文本流畅度更高；\u003C\u002Fli>\n\u003Cli>缺点是内存、算力开销大，线上推理极少使用。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw15.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xpIhMTelP0%2F0WIGWUvO8qiW%2BshI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"hugging-face-与-transformers-库\">\u003Ca class=\"anchor\" href=\"#hugging-face-与-transformers-库\" aria-label=\"跳转到标题：Hugging Face 与 Transformers 库\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Hugging Face 与 Transformers 库\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw16.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SrHJsrDjJoXSg8g8SKv6e6HWQR4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Ccode>config.json\u003C\u002Fcode>：存储模型全局超参数（隐藏层维度、头数、MLP 尺寸等）；\u003C\u002Fli>\n\u003Cli>Safetensors 以安全、高效的方式保存模型张量数据，通过参数名称映射到模型结构中的对应权重。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw17.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y4OYR5zR4QmRO9TxY77o4X%2Bu1yE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_79_zw18.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MedvDk%2B5u7LV7hAsWOGhpBu%2FuS0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"课程作业提示\">\u003Ca class=\"anchor\" href=\"#课程作业提示\" aria-label=\"跳转到标题：课程作业提示\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>课程作业提示\u003C\u002Fh2>\u003Col>\n\u003Cli>仔细阅读作业 README，所有实现步骤、约束均有完整说明；\u003C\u002Fli>\n\u003Cli>作业难度较高，建议尽早动手实践，疑问可在训练营官方社群咨询；\u003C\u002Fli>\n\u003Cli>后续课程会安排专项作业答疑与完整项目讲解。\u003C\u002Fli>\n\u003C\u002Fol>\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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j5LLQWJGl7lfBWs0J77BjrIj89Q%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=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5J3BUGj%2FbNiiRCskyxeUoMVLor4%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,133,136,138,141,143,145,149,152,155,158,161,164,166,169,171,174,177,180,183,186,188,190,193],{"id":129,"text":130,"level":110},"因果语言模型-文本生成","因果语言模型：文本生成",{"id":132,"text":132,"level":9},"分类任务到文本生成转换",{"id":134,"text":135,"level":9},"逐-token-多轮生成流程","逐 Token 多轮生成流程",{"id":137,"text":137,"level":9},"原生推理性能缺陷",{"id":139,"text":140,"level":110},"kv-cache-推理核心加速机制","KV Cache 推理核心加速机制",{"id":142,"text":142,"level":9},"核心原理",{"id":144,"text":144,"level":9},"两大推理阶段划分",{"id":146,"text":147,"level":148},"1-prefill-阶段","1.Prefill 阶段",4,{"id":150,"text":151,"level":148},"2-decode-阶段","2.Decode 阶段",{"id":153,"text":154,"level":110},"ai-对话系统实现方案","AI 对话系统实现方案",{"id":156,"text":157,"level":9},"对话本质-带角色标识的文本续写","对话本质：带角色标识的文本续写",{"id":159,"text":160,"level":9},"messages-对话数据结构","Messages 对话数据结构",{"id":162,"text":163,"level":9},"chat-template-对话模板引擎","Chat Template 对话模板引擎",{"id":165,"text":165,"level":110},"生成采样策略详解",{"id":167,"text":168,"level":9},"argmax-采样","Argmax 采样",{"id":170,"text":170,"level":9},"随机基础采样",{"id":172,"text":173,"level":9},"top-k-采样","Top-K 采样",{"id":175,"text":176,"level":9},"top-p-核采样","Top-P 核采样",{"id":178,"text":179,"level":9},"temperature-温度系数","Temperature 温度系数",{"id":181,"text":182,"level":9},"beam-search","Beam Search",{"id":184,"text":185,"level":110},"hugging-face-与-transformers-库","Hugging Face 与 Transformers 库",{"id":187,"text":187,"level":110},"课程作业提示",{"id":189,"text":189,"level":110},"加入训练营",{"id":191,"text":192,"level":148},"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：",{"id":194,"text":194,"level":110},"关注与交流"]