[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-88":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_88_fm.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=nJx2fxXjATXWHpXcnY3%2BLXTwdfM%3D","大模型推理,KV Cache,量化,分布式推理,FlashAttention","讲解大模型推理核心痛点，覆盖 FP8\u002FINT4 量化、GQA\u002FMoE、FlashAttention、投机采样，详解张量 \u002F 流水线并行等分布式推理混合部署方案。 ","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营推理引擎 4：《大模型计算优化与分布式推理》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MhUpumKcy9%2FUeHLrwnwJiFHiTME%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 31 日，2026 夏季 InfiniTensor 训练营推理引擎方向第四课《大模型计算优化与分布式推理》开讲。\r\n前三课依次讲解 AI 模型导论、Transformer 底层结构、单卡 KV Cache 对话推理完整流程，铺垫基础生成逻辑与 Prefill \u002F Decode 双阶段机制。\u003C\u002Fspan>\r\n\r\n本节课解决超大模型单卡放不下、推理速度慢的工程痛点，聚焦大模型推理系统的核心优化技术与分布式部署方案。\r\n\r\n## 回顾：推理系统的挑战\r\n### 文本生成机制\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jr5XVru1OB%2FU3uGlnuinHgnXszQ%3D)\r\n\r\n1.  Prefill 阶段：填充 KV Cache，避免重复计算\r\n2. Decode 阶段：自回归生成 Token 直至停止符\r\n3. Template 机制：通过身份标识实现多轮 AI 对话\r\n\r\n### Scaling Law 带来的困境\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g9XUSfWNTS5BX2IyGCx9UO3%2B5Q4%3D)\r\n\r\n1. 模型尺寸激增：FP16 下 1B 参数 ≈ 2GB 显存，实际推理还需额外考虑 KV Cache，2.8T 参数需海量显卡\r\n2. KV Cache 膨胀：长上下文场景下 Cache 大小可能超越模型本身\r\n3. 计算延迟上升：算子数量与张量规模同步增长\r\n4. 推理优化的“不可能三角”\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KBi1hzX61XsTgDXaZwA6Uv8eizw%3D)\r\n\r\n\r\n## 数据类型优化：低精度推理基础\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2X5uJ7JzQvDrE28a3Bnn2MTVtpk%3D)\r\n\r\n1.**浮点数类型详解**\r\n- FP32：1 符号位+ 8 指数位 + 23 小数位\r\n- FP16 vs BF16：同为 16bit，指数位与小数位分配不同\r\n- TF32（NVIDIA 特有）：采用 FP32 的 8 位指数位和 10 位尾数位\r\n\r\n2.**低精度推理**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z8ZWjT0Jn1b3ekYcN28mmlCI1PI%3D)\r\n\r\n- 权重存储相比 FP32 减半，同时降低部分激活与缓存存储开销\r\n- 硬件加速支持（Tensor Core 等）\r\n- 敏感算子处理：`rms_norm`、`softmax` 等自动回退 FP32 计算\r\n\r\n## 量化推理\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5nPLARI3oyrnBU3WQ28GepYGElw%3D)\r\n\r\n1.**量化基本原理**\r\n- 浮点到整数的映射：利用参数正态\u002F长尾分布特性\r\n- W\u002FA 命名规范：W4A16 = 权重 INT4 + 激活值 FP16\r\n\r\n2.**两大量化范式**\r\n- PTQ（训练后量化）：主流方案，适用于已训好模型\r\n- QAT（训练时量化）：训练中同步量化\r\n\r\n3.**量化示例（Qwen3-235B-A22B）**\r\n- **FP8 -E4M3**：权重 FP8 + Scale 还原，支持 W8A16\u002FW8A8\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=C9CGyVWcU5V44w4r%2BHeaEIR1qpI%3D)\r\n\r\n\r\n- **GPTQ -INT4**：每行8个 INT4 组成 INT32，Group=128 共享 scale\u002Fqzero，支持重排\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NuEztaOB%2BoUEtOcYTQLb4jBhc9M%3D)\r\n\r\n- **AWQ -INT4**：按输出维度量化，原理类似 GPTQ\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e2MfsDorOkZJdFCEU8nYdBz8Rp4%3D)\r\n\r\n- **Quark-MXFP4**：4bit浮点+E8M0 Scale，极小表示范围\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=m1cxEVW%2BCfmFCBh1SNPiDh8iT9E%3D)\r\n\r\n## 模型结构优化\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z%2B6Ldr2wSt%2FrZYhtJv4BDA20Gqs%3D)\r\n\r\n1.**从 MHA 到 MQA、GQA**\r\n- MQA：多个 Query head 共享同一个 KV head，从而减少 KV Cache 存储量\r\n- GQA（分组查询注意力）：折中方案，当前主流\r\n\r\n2.**MoE（混合专家模型）**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw13.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=drhQEtWcJ1MbEtiCTHoRRffGH4M%3D)\r\n\r\n- 稠密 MLP 拆分为多个专家子网络\r\n- 动态路由选择 Top-K 专家激活\r\n- 效果：保持更大模型容量的同时，仅激活部分专家，从而降低单 Token 推理计算量\r\n\r\n## 计算层优化\r\n1.**融合算子**\r\n- QKV 矩阵乘融合：减少访存与 Kernel 发射\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw14.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AiE0gPS9AI8mQcgM93IENqNTvBs%3D)\r\n\r\n2.**Flash Attention**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw15.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5AlMCfW%2Bxd2g3AiOtrtC4AVxBBA%3D)\r\n\r\n- 问题：标准 Attention 虽然可以分块计算，但通常需要显式存储 Attention Score 矩阵，导致大量 HBM 读写。FlashAttention 利用 Online Softmax 实现 IO-aware 分块计算，避免生成完整 Attention Matrix\r\n- 解法：Online Softmax 算法实现分块计算\r\n- 收益：数据驻留 SRAM，消除 Score 矩阵反复读写\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw16.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WV6EzhDCFXaGTkfWHGd%2FuollKR8%3D)\r\n\r\n3.**投机采样**\r\n\r\n- 核心思路：小模型生成 + 大模型纠错\r\n- 变体：MTP 模块（DeepSeek）、EAGLE、DSpark 等\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw17.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jX%2FL3aiYqZ5gpB%2BXTznV9r7nnq4%3D)\r\n\r\n\r\n## 分布式推理：突破单机显存瓶颈\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw18.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JTucgeZbVFsb3zvRpur4jy%2FnlpA%3D)\r\n\r\n### 并行策略概览\r\n1. 数据并行：提升吞吐，不解决单模型过大问题\r\n2. 模型并行：流水线并行（层间） + 张量并行（层内）\r\n\r\n### 张量并行详解\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw19.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XX7BOSTIynh1wAYoj3lfU89O3Yo%3D)\r\n\r\n1. **张量状态**：Replication \u002F Partition \u002F Partial Sum\r\n2.**分布式矩阵乘**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw20.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e%2BVSiXx%2FCRBsEs8qnfMWX3f5AIQ%3D)\r\n\r\n3. **MLP 层切分**：gate_proj 和 up_proj 通常按 intermediate_size 维度切分，down_proj 按对应输入维度切分，并通过 AllReduce 汇总结果\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw21.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RdtCz%2FJ2QYhRgkwODZiTZEIZ6%2BA%3D)\r\n\r\n\r\n4. **Attention 层切分**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw22.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IvTmxYgGOU%2B4rzPvOy%2FwzYenmBs%3D)\r\n\r\n\r\n\r\n### 流水线并行（PP）与混合部署\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw23.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Gq868OGQvQnLgAN%2FaAe2luzamMk%3D)\r\n\r\n1. 推理场景特点：主要用于解决跨机显存\r\n2. **主流混合并行策略**：节点内 TP + 节点间 PP\r\n3. 跨机通信优化：send\u002Frecv 切分传输 + 节点内 AllGather，降低带宽压力\r\n\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\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-87","study-87","2026 夏季 InfiniTensor 训练营 CUDA 编程 6：《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 1》",21,"\u002Fstudy-material\u002Fstudy-88","study-88","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-90","study-90","2026 夏季 InfiniTensor 训练营 CUDA 编程 7：《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 2》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw01.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MhUpumKcy9%2FUeHLrwnwJiFHiTME%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 31 日，2026 夏季 InfiniTensor 训练营推理引擎方向第四课《大模型计算优化与分布式推理》开讲。\u003Cbr>前三课依次讲解 AI 模型导论、Transformer 底层结构、单卡 KV Cache 对话推理完整流程，铺垫基础生成逻辑与 Prefill \u002F Decode 双阶段机制。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课解决超大模型单卡放不下、推理速度慢的工程痛点，聚焦大模型推理系统的核心优化技术与分布式部署方案。\u003C\u002Fp>\n\u003Ch2 id=\"回顾-推理系统的挑战\">\u003Ca class=\"anchor\" href=\"#回顾-推理系统的挑战\" aria-label=\"跳转到标题：回顾：推理系统的挑战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>回顾：推理系统的挑战\u003C\u002Fh2>\u003Ch3 id=\"文本生成机制\">\u003Ca class=\"anchor\" href=\"#文本生成机制\" aria-label=\"跳转到标题：文本生成机制\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>文本生成机制\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw02.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Jr5XVru1OB%2FU3uGlnuinHgnXszQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>Prefill 阶段：填充 KV Cache，避免重复计算\u003C\u002Fli>\n\u003Cli>Decode 阶段：自回归生成 Token 直至停止符\u003C\u002Fli>\n\u003Cli>Template 机制：通过身份标识实现多轮 AI 对话\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"scaling-law-带来的困境\">\u003Ca class=\"anchor\" href=\"#scaling-law-带来的困境\" aria-label=\"跳转到标题：Scaling Law 带来的困境\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Scaling Law 带来的困境\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw03.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g9XUSfWNTS5BX2IyGCx9UO3%2B5Q4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>模型尺寸激增：FP16 下 1B 参数 ≈ 2GB 显存，实际推理还需额外考虑 KV Cache，2.8T 参数需海量显卡\u003C\u002Fli>\n\u003Cli>KV Cache 膨胀：长上下文场景下 Cache 大小可能超越模型本身\u003C\u002Fli>\n\u003Cli>计算延迟上升：算子数量与张量规模同步增长\u003C\u002Fli>\n\u003Cli>推理优化的“不可能三角”\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw04.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=KBi1hzX61XsTgDXaZwA6Uv8eizw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"数据类型优化-低精度推理基础\">\u003Ca class=\"anchor\" href=\"#数据类型优化-低精度推理基础\" aria-label=\"跳转到标题：数据类型优化：低精度推理基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>数据类型优化：低精度推理基础\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw05.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2X5uJ7JzQvDrE28a3Bnn2MTVtpk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>浮点数类型详解\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>FP32：1 符号位+ 8 指数位 + 23 小数位\u003C\u002Fli>\n\u003Cli>FP16 vs BF16：同为 16bit，指数位与小数位分配不同\u003C\u002Fli>\n\u003Cli>TF32（NVIDIA 特有）：采用 FP32 的 8 位指数位和 10 位尾数位\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>低精度推理\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw06.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z8ZWjT0Jn1b3ekYcN28mmlCI1PI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>权重存储相比 FP32 减半，同时降低部分激活与缓存存储开销\u003C\u002Fli>\n\u003Cli>硬件加速支持（Tensor Core 等）\u003C\u002Fli>\n\u003Cli>敏感算子处理：\u003Ccode>rms_norm\u003C\u002Fcode>、\u003Ccode>softmax\u003C\u002Fcode> 等自动回退 FP32 计算\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"量化推理\">\u003Ca class=\"anchor\" href=\"#量化推理\" aria-label=\"跳转到标题：量化推理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化推理\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw07.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5nPLARI3oyrnBU3WQ28GepYGElw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>量化基本原理\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>浮点到整数的映射：利用参数正态\u002F长尾分布特性\u003C\u002Fli>\n\u003Cli>W\u002FA 命名规范：W4A16 = 权重 INT4 + 激活值 FP16\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>两大量化范式\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>PTQ（训练后量化）：主流方案，适用于已训好模型\u003C\u002Fli>\n\u003Cli>QAT（训练时量化）：训练中同步量化\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>3.\u003Cstrong>量化示例（Qwen3-235B-A22B）\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>FP8 -E4M3\u003C\u002Fstrong>：权重 FP8 + Scale 还原，支持 W8A16\u002FW8A8\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw08.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=C9CGyVWcU5V44w4r%2BHeaEIR1qpI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>GPTQ -INT4\u003C\u002Fstrong>：每行8个 INT4 组成 INT32，Group=128 共享 scale\u002Fqzero，支持重排\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw09.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NuEztaOB%2BoUEtOcYTQLb4jBhc9M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AWQ -INT4\u003C\u002Fstrong>：按输出维度量化，原理类似 GPTQ\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw10.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e2MfsDorOkZJdFCEU8nYdBz8Rp4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Quark-MXFP4\u003C\u002Fstrong>：4bit浮点+E8M0 Scale，极小表示范围\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw11.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=m1cxEVW%2BCfmFCBh1SNPiDh8iT9E%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"模型结构优化\">\u003Ca class=\"anchor\" href=\"#模型结构优化\" aria-label=\"跳转到标题：模型结构优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>模型结构优化\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw12.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=z%2B6Ldr2wSt%2FrZYhtJv4BDA20Gqs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>从 MHA 到 MQA、GQA\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>MQA：多个 Query head 共享同一个 KV head，从而减少 KV Cache 存储量\u003C\u002Fli>\n\u003Cli>GQA（分组查询注意力）：折中方案，当前主流\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>MoE（混合专家模型）\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw13.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=drhQEtWcJ1MbEtiCTHoRRffGH4M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>稠密 MLP 拆分为多个专家子网络\u003C\u002Fli>\n\u003Cli>动态路由选择 Top-K 专家激活\u003C\u002Fli>\n\u003Cli>效果：保持更大模型容量的同时，仅激活部分专家，从而降低单 Token 推理计算量\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"计算层优化\">\u003Ca class=\"anchor\" href=\"#计算层优化\" aria-label=\"跳转到标题：计算层优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算层优化\u003C\u002Fh2>\u003Cp>1.\u003Cstrong>融合算子\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>QKV 矩阵乘融合：减少访存与 Kernel 发射\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw14.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=AiE0gPS9AI8mQcgM93IENqNTvBs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>Flash Attention\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw15.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5AlMCfW%2Bxd2g3AiOtrtC4AVxBBA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>问题：标准 Attention 虽然可以分块计算，但通常需要显式存储 Attention Score 矩阵，导致大量 HBM 读写。FlashAttention 利用 Online Softmax 实现 IO-aware 分块计算，避免生成完整 Attention Matrix\u003C\u002Fli>\n\u003Cli>解法：Online Softmax 算法实现分块计算\u003C\u002Fli>\n\u003Cli>收益：数据驻留 SRAM，消除 Score 矩阵反复读写\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw16.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WV6EzhDCFXaGTkfWHGd%2FuollKR8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>投机采样\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>核心思路：小模型生成 + 大模型纠错\u003C\u002Fli>\n\u003Cli>变体：MTP 模块（DeepSeek）、EAGLE、DSpark 等\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw17.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jX%2FL3aiYqZ5gpB%2BXTznV9r7nnq4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"分布式推理-突破单机显存瓶颈\">\u003Ca class=\"anchor\" href=\"#分布式推理-突破单机显存瓶颈\" aria-label=\"跳转到标题：分布式推理：突破单机显存瓶颈\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>分布式推理：突破单机显存瓶颈\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw18.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JTucgeZbVFsb3zvRpur4jy%2FnlpA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"并行策略概览\">\u003Ca class=\"anchor\" href=\"#并行策略概览\" aria-label=\"跳转到标题：并行策略概览\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>并行策略概览\u003C\u002Fh3>\u003Col>\n\u003Cli>数据并行：提升吞吐，不解决单模型过大问题\u003C\u002Fli>\n\u003Cli>模型并行：流水线并行（层间） + 张量并行（层内）\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"张量并行详解\">\u003Ca class=\"anchor\" href=\"#张量并行详解\" aria-label=\"跳转到标题：张量并行详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>张量并行详解\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw19.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XX7BOSTIynh1wAYoj3lfU89O3Yo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>张量状态\u003C\u002Fstrong>：Replication \u002F Partition \u002F Partial Sum\u003Cbr>2.\u003Cstrong>分布式矩阵乘\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw20.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e%2BVSiXx%2FCRBsEs8qnfMWX3f5AIQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>\u003Cstrong>MLP 层切分\u003C\u002Fstrong>：gate_proj 和 up_proj 通常按 intermediate_size 维度切分，down_proj 按对应输入维度切分，并通过 AllReduce 汇总结果\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw21.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RdtCz%2FJ2QYhRgkwODZiTZEIZ6%2BA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"4\">\n\u003Cli>\u003Cstrong>Attention 层切分\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw22.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IvTmxYgGOU%2B4rzPvOy%2FwzYenmBs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"流水线并行-pp-与混合部署\">\u003Ca class=\"anchor\" href=\"#流水线并行-pp-与混合部署\" aria-label=\"跳转到标题：流水线并行（PP）与混合部署\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>流水线并行（PP）与混合部署\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_88_zw23.jpg?Expires=1792752589&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Gq868OGQvQnLgAN%2FaAe2luzamMk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>推理场景特点：主要用于解决跨机显存\u003C\u002Fli>\n\u003Cli>\u003Cstrong>主流混合并行策略\u003C\u002Fstrong>：节点内 TP + 节点间 PP\u003C\u002Fli>\n\u003Cli>跨机通信优化：send\u002Frecv 切分传输 + 节点内 AllGather，降低带宽压力\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",[128,131,133,136,139,141,143,145,148,150,152,155,157,161],{"id":129,"text":130,"level":110},"回顾-推理系统的挑战","回顾：推理系统的挑战",{"id":132,"text":132,"level":9},"文本生成机制",{"id":134,"text":135,"level":9},"scaling-law-带来的困境","Scaling Law 带来的困境",{"id":137,"text":138,"level":110},"数据类型优化-低精度推理基础","数据类型优化：低精度推理基础",{"id":140,"text":140,"level":110},"量化推理",{"id":142,"text":142,"level":110},"模型结构优化",{"id":144,"text":144,"level":110},"计算层优化",{"id":146,"text":147,"level":110},"分布式推理-突破单机显存瓶颈","分布式推理：突破单机显存瓶颈",{"id":149,"text":149,"level":9},"并行策略概览",{"id":151,"text":151,"level":9},"张量并行详解",{"id":153,"text":154,"level":9},"流水线并行-pp-与混合部署","流水线并行（PP）与混合部署",{"id":156,"text":156,"level":110},"加入训练营",{"id":158,"text":159,"level":160},"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":162,"text":162,"level":110},"关注与交流"]