[{"data":1,"prerenderedAt":211},["ShallowReactive",2],{"content-query-Vvkel5pjKd":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":205,"_id":206,"_source":207,"_file":208,"_stem":209,"_extension":210},"/news/zh/2026-4-30","zh",false,"","MindSpore AKG Meetup北京站圆满举办 | 产学研聚力 共探算子技术创新 共建开源生态","活动线上线下同步开展，汇聚北航计算机学院、昇思社区等超过3000名开发者，聚焦算子智能生成、大模型编译等方向，以产学研协同推动 AKG 技术迭代与开源生态建设。","2026-4-30","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/07/25/199b735845bf4106b44b2035dc97bd39.png","news",{"type":14,"children":15,"toc":198},"root",[16,24,35,41,46,51,58,65,72,77,83,88,93,98,105,110,115,122,127,132,139,144,151,156,161,168,173,180,186,191],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"mindspore-akg-meetup北京站圆满举办-产学研聚力-共探算子技术创新-共建开源生态",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":28},"div",{"style":27},"text-align: center;",[29],{"type":17,"tag":30,"props":31,"children":34},"img",{"src":32,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/1.jpg","display: block;margin: 0 auto;max-width:70%",[],{"type":17,"tag":36,"props":37,"children":38},"p",{},[39],{"type":23,"value":40},"随着大模型代码生成能力不断成熟，高校与厂商纷纷投身 LLM 算子生成技术探索。",{"type":17,"tag":36,"props":42,"children":43},{},[44],{"type":23,"value":45},"4 月 29 日，昇思 MindSpore 社区与北京航空航天大学联合举办 MindSpore AKG Meetup 北京站。活动线上线下同步开展，汇聚北航计算机学院、昇思社区等超过3000名开发者，聚焦算子智能生成、大模型编译等方向，以产学研协同推动 AKG 技术迭代与开源生态建设。",{"type":17,"tag":36,"props":47,"children":48},{},[49],{"type":23,"value":50},"AKG Agent 是本次活动核心技术成果，自 2025 年初布局 Agent 算子生成能力，现已打造自动化 Multi-Agent 生成系统，搭载分层 Skill、自适应并发搜索、AutoResearch 深度搜索等核心技术，昇腾算子生成效果优异：Triton-ascend 78% 达 0.8x torch_npu、66% 达 1.0x torch_npu（Eager 模式）。",{"type":17,"tag":52,"props":53,"children":55},"h2",{"id":54},"_01-开场致辞明确共建方向",[56],{"type":23,"value":57},"01 开场致辞：明确共建方向",{"type":17,"tag":25,"props":59,"children":60},{"style":27},[61],{"type":17,"tag":30,"props":62,"children":64},{"src":63,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/2.jpg",[],{"type":17,"tag":25,"props":66,"children":67},{"style":27},[68],{"type":17,"tag":30,"props":69,"children":71},{"src":70,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/3.jpg",[],{"type":17,"tag":36,"props":73,"children":74},{},[75],{"type":23,"value":76},"北京航空航天大学计算机学院副院长杨海龙致辞，期待以本次活动为纽带深化校企合作，推动科研成果转化，培育算子智能生成领域人才。",{"type":17,"tag":52,"props":78,"children":80},{"id":79},"_02-技术报告环节",[81],{"type":23,"value":82},"02 技术报告环节",{"type":17,"tag":36,"props":84,"children":85},{},[86],{"type":23,"value":87},"致辞结束后，活动进入核心技术报告环节。多位专家围绕算子智能生成、多智能体优化、异构计算、稀疏大模型以及AI辅助开发等主题进行了系统分享。",{"type":17,"tag":36,"props":89,"children":90},{},[91],{"type":23,"value":92},"Part 1 ：算子智能生成与多智能体优化\nMindSpore AKG SIG 技术专家易衍智分享《基于 Agent 的算子自动生成简介与 AKG Agent》，讲解多角色架构、模块化 Skill 系统，以及 Sketch 并发搜索与 AutoResearch 优化框架。",{"type":17,"tag":36,"props":94,"children":95},{},[96],{"type":23,"value":97},"实际应用中，该方案可降低 20% Token 消耗、提升 10% 算子性能、缩短 30% 整体耗时；矩阵计算、混合分组直方图等算子耗时从 2193us 降至 168us、8245us 降至 606us，KernelBench Level1 算子通过率 100%。",{"type":17,"tag":25,"props":99,"children":100},{"style":27},[101],{"type":17,"tag":30,"props":102,"children":104},{"src":103,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/4.jpg",[],{"type":17,"tag":36,"props":106,"children":107},{},[108],{"type":23,"value":109},"北京航空航天大学游心老师分享《基于多智能体协同优化的高性能算子自动生成系统》，针对算力供需失衡问题，提出细粒度元提示、测试驱动自修复、评测驱动协同优化三大技术。",{"type":17,"tag":36,"props":111,"children":112},{},[113],{"type":23,"value":114},"实测效果：GPU 算子生成 Pass@1 从 88% 升至 95%，NPU 从 69% 升至 71%；CPU 平均加速 2.81 倍、GPU 平均加速 2.30 倍，最高达 10.95 倍，可稳定运行 Attention 等复杂算子。",{"type":17,"tag":25,"props":116,"children":117},{"style":27},[118],{"type":17,"tag":30,"props":119,"children":121},{"src":120,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/5.jpg",[],{"type":17,"tag":36,"props":123,"children":124},{},[125],{"type":23,"value":126},"Part 2 ：异构计算与稀疏大模型优化\n北京航空航天大学解晨浩老师分享《面向异构超算的多维分块稀疏矩阵研究及应用》，提出多维分块稀疏矩阵计算模型。",{"type":17,"tag":36,"props":128,"children":129},{},[130],{"type":23,"value":131},"实际效果：油气藏数值模拟场景中，100 节点较 50 节点实现近 2 倍加速；DiT 模型推理中，RT-Lynx 稀疏算子最高加速 1.79 倍，显著提升异构超算与稀疏计算效率。",{"type":17,"tag":25,"props":133,"children":134},{"style":27},[135],{"type":17,"tag":30,"props":136,"children":138},{"src":137,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/6.jpg",[],{"type":17,"tag":36,"props":140,"children":141},{},[142],{"type":23,"value":143},"北京航空航天大学孙庆骁老师，分享《面向稀疏大模型推理的算子实现及编译优化研究介绍》，核心如下：\n1.痛点直击：针对稀疏大模型推理瓶颈，明确算子开发与编译优化需求。\n2.优化手段：讲解编译优化、内存对齐等提升推理效率的方法。\n3.解决方案：提供大模型轻量化部署与高效推理的完整方案。",{"type":17,"tag":25,"props":145,"children":146},{"style":27},[147],{"type":17,"tag":30,"props":148,"children":150},{"src":149,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/7.jpg",[],{"type":17,"tag":36,"props":152,"children":153},{},[154],{"type":23,"value":155},"Part 3 ：开源生态与AI辅助开发\nCANN 开源技术专家张雨晨分享《CANN 社区及 CANNBot 介绍》，介绍 CANN 社区生态与 CANNBot 三层架构、全流程自动化工作流。",{"type":17,"tag":36,"props":157,"children":158},{},[159],{"type":23,"value":160},"实际效果：CANNBot 将 Ascend C 算子开发周期从天级压缩至小时级，Abs 算子仅需 1 小时即可完成全流程开发，大幅降低开发门槛。",{"type":17,"tag":25,"props":162,"children":163},{"style":27},[164],{"type":17,"tag":30,"props":165,"children":167},{"src":166,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/8.jpg",[],{"type":17,"tag":36,"props":169,"children":170},{},[171],{"type":23,"value":172},"操作系统技术专家李屹，压轴分享《从测试、代码迁移到端到端功能实现 —AI 辅助开发的思考与程序员的未来》，核心如下：\n1.实践历程：分享团队AI开发三阶段及效率提升成效。\n2.AI局限：指出大模型能力边界，难以兼顾架构优美与需求深层解读。\n3.未来启示：AI替代重复性劳动，程序员核心竞争力在业务洞察与创新。",{"type":17,"tag":25,"props":174,"children":175},{"style":27},[176],{"type":17,"tag":30,"props":177,"children":179},{"src":178,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/9.jpg",[],{"type":17,"tag":52,"props":181,"children":183},{"id":182},"_03-总结与展望",[184],{"type":23,"value":185},"03 总结与展望",{"type":17,"tag":36,"props":187,"children":188},{},[189],{"type":23,"value":190},"本次MindSpore AKG Meetup北京站搭建了产学研三方交流平台，普及前沿技术、凝聚开源力量、深化校企融合。未来，昇思MindSpore AKG将持续落地多地Meetup活动，深耕算力技术创新，汇聚开发者智慧，打造开放共赢的AI开源生态。",{"type":17,"tag":25,"props":192,"children":193},{"style":27},[194],{"type":17,"tag":30,"props":195,"children":197},{"src":196,"style":33,"alt":7},"/category/information/news/banner/2026-4-30/10.jpg",[],{"title":7,"searchDepth":199,"depth":199,"links":200},4,[201,203,204],{"id":54,"depth":202,"text":57},2,{"id":79,"depth":202,"text":82},{"id":182,"depth":202,"text":185},"markdown","content:news:zh:2026-4-30.md","content","news/zh/2026-4-30.md","news/zh/2026-4-30","md",1778320583578]