[{"data":1,"prerenderedAt":232},["ShallowReactive",2],{"content-query-8t8YT1hHMX":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":226,"_id":227,"_source":228,"_file":229,"_stem":230,"_extension":231},"\u002Fnews\u002Fzh\u002F2026-7-10","zh",false,"","昇思MindSpore社区 2026 年 6 月份运作报告","昇思MindSpore开源社区持续深耕技术突破与生态共建，在AI编译技术、分布式训练、开发者效能等方面取得多项重要成果。","2026-7-10","https:\u002F\u002Fobs-mindspore-file.obs.cn-north-4.myhuaweicloud.com\u002Ffile\u002F2025\u002F07\u002F25\u002F199b735845bf4106b44b2035dc97bd39.png","news",{"type":14,"children":15,"toc":218},"root",[16,28,36,42,47,52,57,63,68,83,90,95,108,114,121,126,133,138,144,149,156,162,167,174,179,185,190,197,202,208,213],{"type":17,"tag":18,"props":19,"children":21},"element","div",{"style":20},"text-align: center;",[22],{"type":17,"tag":23,"props":24,"children":27},"img",{"src":25,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F1.jpg","display: block;margin: 0 auto;max-width:60%",[],{"type":17,"tag":29,"props":30,"children":32},"h1",{"id":31},"一概述",[33],{"type":34,"value":35},"text","一、概述",{"type":17,"tag":37,"props":38,"children":39},"p",{},[40],{"type":34,"value":41},"2026年6月，昇思MindSpore开源社区持续深耕技术突破与生态共建，在AI编译技术、分布式训练、开发者效能等方面取得多项重要成果。",{"type":17,"tag":37,"props":43,"children":44},{},[45],{"type":34,"value":46},"在技术分享方面，业界首个动态模型亲和的“AI实时编译器”MindSpore DVM正式开源，编译时间最高提速5个数量级，并同步推出动态图无图融合能力，为运行时实时融合提供新路径；MindSpore HyperParallel深度优化盘古Muon优化器，在千亿MoE模型上实现6倍性能跃升，同时基于DTensor的AI辅助编程体系将分布式算子开发效率提升83%。",{"type":17,"tag":37,"props":48,"children":49},{},[50],{"type":34,"value":51},"在社区运作方面，社区治理与协作体系高效运转，技术贡献与活跃度持续攀升，昇思正携手开发者共筑全场景AI开源创新生态。",{"type":17,"tag":37,"props":53,"children":54},{},[55],{"type":34,"value":56},"以下是2026年6月份昇思MindSpore社区进展详细报告。",{"type":17,"tag":29,"props":58,"children":60},{"id":59},"二社区运作",[61],{"type":34,"value":62},"二、社区运作",{"type":17,"tag":37,"props":64,"children":65},{},[66],{"type":34,"value":67},"截至6月底，MindSpore 开源社区稳健运营、生态持续扩容，始终坚守开放协作、技术普惠、共建共治的核心原则，聚焦 AI 框架技术迭代、生态协同与开发者生态建设，社区组织架构高效运转、技术贡献与活跃度双提升，筑牢 AI 开源生态的坚实底座，推动全场景 AI 框架技术与产业应用深度融合。",{"type":17,"tag":69,"props":70,"children":71},"ul",{},[72,78],{"type":17,"tag":73,"props":74,"children":75},"li",{},[76],{"type":34,"value":77},"社区单位会员达 356 家，累计贡献者5.4万+；",{"type":17,"tag":73,"props":79,"children":80},{},[81],{"type":34,"value":82},"社区累计产生ISSUE共91.8K、PR共131.8K。",{"type":17,"tag":18,"props":84,"children":85},{"style":20},[86],{"type":17,"tag":23,"props":87,"children":89},{"src":88,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F2.jpg",[],{"type":17,"tag":37,"props":91,"children":92},{},[93],{"type":34,"value":94},"6月份社区治理体系高效运转，开源发展委员会、技术委员会常态化推进决策与技术统筹，各 SIG 小组协同发力，技术贡献与社区活跃度稳步攀升。",{"type":17,"tag":69,"props":96,"children":97},{},[98,103],{"type":17,"tag":73,"props":99,"children":100},{},[101],{"type":34,"value":102},"6月提交issue 670 个，合入PR 1073 个；",{"type":17,"tag":73,"props":104,"children":105},{},[106],{"type":34,"value":107},"新增代码行数超 126万行，SIG 专项会议累计召开 5场。",{"type":17,"tag":29,"props":109,"children":111},{"id":110},"三技术分享",[112],{"type":34,"value":113},"三、技术分享",{"type":17,"tag":115,"props":116,"children":118},"h2",{"id":117},"_01-业界首个动态模型亲和的ai实时编译器mindspore-dvm强力推荐",[119],{"type":34,"value":120},"01 业界首个动态模型亲和的“AI实时编译器”—MindSpore DVM，强力推荐",{"type":17,"tag":37,"props":122,"children":123},{},[124],{"type":34,"value":125},"MindSpore联合港科广、湖南大学推出业界首个动态模型亲和的“AI实时编译器”DVM。传统编译器依赖离线shape推导，难以应对动态模型；运行时编译链路过长。DVM不生成机器码，而是生成tile级字节码，由NPU侧虚拟机解释执行，将编译开销隐藏在硬件流水线中。支持pattern与streaming融合，动态图中根据真实执行路径即时决策。",{"type":17,"tag":18,"props":127,"children":128},{"style":20},[129],{"type":17,"tag":23,"props":130,"children":132},{"src":131,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F3.jpg",[],{"type":17,"tag":37,"props":134,"children":135},{},[136],{"type":34,"value":137},"LayerNorm上相较TorchInductor加速11.77倍，Qwen3-14B微调提升1.07~1.21倍，编译时间最高提速5个数量级。已开源，支持MindSpore与PyTorch后端。",{"type":17,"tag":115,"props":139,"children":141},{"id":140},"_02-mindspore-推出动态图无图融合能力",[142],{"type":34,"value":143},"02 MindSpore 推出动态图无图融合能力",{"type":17,"tag":37,"props":145,"children":146},{},[147],{"type":34,"value":148},"MindSpore推出动态图无图融合能力，在运行时沿真实算子流增量捕捉与轻量构图，由底层DVM实时编译生成融合Kernel。相比静态图离线编译，该方案不依赖图缓存，可适应动态shape变化。当前支持Elemwise相互融合、Reduce前向融合及MatMul+Elemwise等CV融合扩展，整网端到端性能收益约1%~5%。通过环境变量MS_DEV_PYNATIVE_FUSION_FLAGS开启，代码已开源。",{"type":17,"tag":18,"props":150,"children":151},{"style":20},[152],{"type":17,"tag":23,"props":153,"children":155},{"src":154,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F4.jpg",[],{"type":17,"tag":115,"props":157,"children":159},{"id":158},"_03-hdc技术速递-mindspore-hyperparallel助力盘古muon优化器实现6倍性能跃升",[160],{"type":34,"value":161},"03 HDC技术速递 | MindSpore HyperParallel助力盘古Muon优化器实现6倍性能跃升",{"type":17,"tag":37,"props":163,"children":164},{},[165],{"type":34,"value":166},"MindSpore HyperParallel基于FSDP实现Muon优化器深度优化，在昇腾512卡集群中，千亿参数MoE模型单次计算耗时由2700ms降至450ms，性能提升6倍。",{"type":17,"tag":18,"props":168,"children":169},{"style":20},[170],{"type":17,"tag":23,"props":171,"children":173},{"src":172,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F5.jpg",[],{"type":17,"tag":37,"props":175,"children":176},{},[177],{"type":34,"value":178},"方案针对三大挑战：全矩阵收集通信开销大、Newton-Schulz迭代Host调度开销高、DP副本间冗余计算严重。团队提出AllGather+DP去冗余方案，计算耗时从485ms收敛至129ms；对高维专家参数按Batch维度独立计算，通信清零、耗时骤降77.5%；通过矩阵融合将小矩阵重组为批量张量，74个参数从74ms降至2.4ms；HSDP场景下副本间分组计算，性能再提升1倍。",{"type":17,"tag":115,"props":180,"children":182},{"id":181},"_04-mindspore-hyperparallelai辅助编程助力分布式算子开发效率提升",[183],{"type":34,"value":184},"04 MindSpore HyperParallel：AI辅助编程，助力分布式算子开发效率提升",{"type":17,"tag":37,"props":186,"children":187},{},[188],{"type":34,"value":189},"MindSpore HyperParallel构建了基于DTensor的AI协作开发体系，通过.agent\u002F目录将专家经验固化为规则与技能资产。",{"type":17,"tag":18,"props":191,"children":192},{"style":20},[193],{"type":17,"tag":23,"props":194,"children":196},{"src":195,"style":26,"alt":7},"\u002Fcategory\u002Finformation\u002Fnews\u002Fbanner\u002F2026-7-10\u002F6.jpg",[],{"type":17,"tag":37,"props":198,"children":199},{},[200],{"type":34,"value":201},"针对分布式算子开发全流程，团队沉淀四个核心Skill：dist-op-analysis进行方案推导与人工确认，dist-op-dev实现自主编码测试，autogit规范提交，gate-doctor自动化门禁修复。实战中tril算子从需求到PR仅用1.5小时，人工操作不足10分钟，较此前约9小时提效83%。该体系在生成阶段即拦截高风险问题，专家经验得以沉淀，不复随人员流失。",{"type":17,"tag":29,"props":203,"children":205},{"id":204},"四感谢每一位朋友开发者的支持",[206],{"type":34,"value":207},"四、感谢每一位朋友、开发者的支持",{"type":17,"tag":37,"props":209,"children":210},{},[211],{"type":34,"value":212},"在此感谢社区伙伴们、可爱的小孢子们以及昇思MindSpore SIG组成员们，因为大家的共同努力及辛勤奉献，昇思MindSpore才能不断成长与发展！同时我们对可能出现的不完善之处向您表示诚挚的歉意，并衷心感谢您的理解与支持。",{"type":17,"tag":37,"props":214,"children":215},{},[216],{"type":34,"value":217},"未来，昇思MindSpore AI框架将持续致力于打造人工智能创新之源，凝聚产业力量，扎根AI根技术，使能大模型与科学智能，成为AI创新的首选框架。",{"title":7,"searchDepth":219,"depth":219,"links":220},4,[221,223,224,225],{"id":117,"depth":222,"text":120},2,{"id":140,"depth":222,"text":143},{"id":158,"depth":222,"text":161},{"id":181,"depth":222,"text":184},"markdown","content:news:zh:2026-7-10.md","content","news\u002Fzh\u002F2026-7-10.md","news\u002Fzh\u002F2026-7-10","md",1784928502040]