[{"data":1,"prerenderedAt":515},["ShallowReactive",2],{"content-query-Y4pjySrypF":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":509,"_id":510,"_source":511,"_file":512,"_stem":513,"_extension":514},"/news/zh/2026-4-9","zh",false,"","昇思MindSpore2026年第一季度运作报告","昇思MindSpore开源社区以技术创新为驱动，持续扩大生态影响力，在框架升级、开发者生态建设、行业应用落地等方面取得显著进展。","2026-4-9","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/07/25/199b735845bf4106b44b2035dc97bd39.png","news",{"type":14,"children":15,"toc":491},"root",[16,24,30,36,41,46,51,56,61,66,71,76,87,92,97,102,107,112,117,122,129,134,139,146,151,156,163,168,173,178,183,190,195,200,205,212,217,222,229,234,241,248,253,258,263,268,275,280,285,290,295,300,305,310,315,322,327,332,339,344,349,354,361,366,372,377,382,387,394,400,405,410,416,421,426,431,437,442,447,452,457,462,475,481,486],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"昇思mindspore2026年第一季度运作报告",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":28},"h2",{"id":27},"概述",[29],{"type":23,"value":27},{"type":17,"tag":31,"props":32,"children":33},"p",{},[34],{"type":23,"value":35},"2026年第一季度，昇思MindSpore开源社区以技术创新为驱动，持续扩大生态影响力，在框架升级、开发者生态建设、行业应用落地等方面取得显著进展。",{"type":17,"tag":31,"props":37,"children":38},{},[39],{"type":23,"value":40},"在技术突破方面，本季度重磅推出MindSpore 2.8版本，核心带来为超节点设计的HyperParallel架构；同时与SGLang社区深度合作，正式支持MindSpore后端，为大模型服务化提供高性能解决方案。",{"type":17,"tag":31,"props":42,"children":43},{},[44],{"type":23,"value":45},"在社区合作方面，助力智谱训练并开源首个自主算力底座的SOTA图像生成模型GLM-Image；联合中国电信AI研究院开源首个全自主创新千亿参数MoE大模型TeleChat3-105B；携手浙江大学张强团队发布知识图谱驱动的SciToolAgent，相关成果登上《Nature Computational Science》。",{"type":17,"tag":31,"props":47,"children":48},{},[49],{"type":23,"value":50},"在社区运作方面，社区通过举办“万码奔腾贺新春”活动开放稀缺算力，并同时开展《开源之夏》栏目共吸引103位学子参与，《昇思同路人》持续传递开发者故事，《创新训练营》优秀案例不断涌现。从大模型全流程课程到量子计算实战，昇思MindSpore正携手全球开发者，共同打造人工智能创新之源！",{"type":17,"tag":31,"props":52,"children":53},{},[54],{"type":23,"value":55},"以下是2026年第一季度昇思MindSpore社区进展的详细报告。",{"type":17,"tag":25,"props":57,"children":59},{"id":58},"社区运作",[60],{"type":23,"value":58},{"type":17,"tag":31,"props":62,"children":63},{},[64],{"type":23,"value":65},"截至3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":31,"props":67,"children":68},{},[69],{"type":23,"value":70},"社区单位会员达 354 家，累计贡献者5.2万+；",{"type":17,"tag":31,"props":72,"children":73},{},[74],{"type":23,"value":75},"社区累计产生ISSUE共89.7K、PR共115.2K。",{"type":17,"tag":77,"props":78,"children":80},"div",{"style":79},"text-align: center;",[81],{"type":17,"tag":82,"props":83,"children":86},"img",{"src":84,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/1.jpg","display: block;margin: 0 auto;max-width:70%",[],{"type":17,"tag":31,"props":88,"children":89},{},[90],{"type":23,"value":91},"本季度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。技术委员会开展 2 次 TC 例会，统筹技术规划与版本演进。",{"type":17,"tag":31,"props":93,"children":94},{},[95],{"type":23,"value":96},"社区 6 个活跃 SIG 小组协同发力，78 名 SIG 成员深度参与贡献；",{"type":17,"tag":31,"props":98,"children":99},{},[100],{"type":23,"value":101},"累计提交 issue 310 个，合入 PR 1102 个；",{"type":17,"tag":31,"props":103,"children":104},{},[105],{"type":23,"value":106},"新增代码行数超 45.9 万行，SIG 专项会议累计召开 24 场。",{"type":17,"tag":25,"props":108,"children":110},{"id":109},"社区大事件",[111],{"type":23,"value":109},{"type":17,"tag":31,"props":113,"children":114},{},[115],{"type":23,"value":116},"1、智谱联合昇腾+昇思，开源首个自主创新算力底座训练的多模态SOTA模型",{"type":17,"tag":31,"props":118,"children":119},{},[120],{"type":23,"value":121},"智谱正式发布图像生成模型GLM-Image，采用「自回归+扩散解码器」混合架构，是首个开源的工业级离散自回归图像生成模型，在文字渲染榜单CVTG-2K和LongText-Bench上达到开源SOTA水平。模型基于昇腾NPU和昇思MindSpore框架完成全流程训练，验证了自主创新算力底座也能训练出SOTA模型。昇思通过动态图多级流水下发、多流并行执行和高性能融合算子等优化，实现端到端训练性能提升。",{"type":17,"tag":77,"props":123,"children":124},{"style":79},[125],{"type":17,"tag":82,"props":126,"children":128},{"src":127,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/2.jpg",[],{"type":17,"tag":31,"props":130,"children":131},{},[132],{"type":23,"value":133},"2、国内首个全自主创新千亿参数细粒度MoE语义大模型开源",{"type":17,"tag":31,"props":135,"children":136},{},[137],{"type":23,"value":138},"TeleChat3-105B-A4.7-Thinking是 TeleChat系列国内首个开源的全自主创新千亿参数细粒度MoE语义大模型，由中国电信人工智能研究院（TeleAI）研发训练，在问答、写作、数学、代码、Agent等多维度，特别在代码能力、复杂任务通用问答、细粒度MoE等方面效果显著。同时采用创新训练方式，加快模型训练初期收敛速度，增强训练稳定性。",{"type":17,"tag":77,"props":140,"children":141},{"style":79},[142],{"type":17,"tag":82,"props":143,"children":145},{"src":144,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/3.jpg",[],{"type":17,"tag":31,"props":147,"children":148},{},[149],{"type":23,"value":150},"3、浙江大学张强团队联手昇思MindSpore，成果登上Nature子刊",{"type":17,"tag":31,"props":152,"children":153},{},[154],{"type":23,"value":155},"张强团队发布知识图谱驱动的SciToolAgent，这是一款面向多学科的通用AI智能体，能在生物、化学、材料等领域自动化调用和管理数百种科学工具。该成果基于昇思MindSpore与昇腾硬件协同创新，相关研究论文发表于《Nature Computational Science》期刊。",{"type":17,"tag":77,"props":157,"children":158},{"style":79},[159],{"type":17,"tag":82,"props":160,"children":162},{"src":161,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/4.jpg",[],{"type":17,"tag":31,"props":164,"children":165},{},[166],{"type":23,"value":167},"4、昇思MindSpore 2.8版本正式发布",{"type":17,"tag":31,"props":169,"children":170},{},[171],{"type":23,"value":172},"1月29日，MindSpore 2.8版本正式发布，核心推出HyperParallel架构，通过HyperShard声明式并行、HyperOffload多级智能卸载和HyperMPMD非规则异构并行三大技术，将超节点视为超级计算机进行编程调度。",{"type":17,"tag":31,"props":174,"children":175},{},[176],{"type":23,"value":177},"基础框架方面，增强动态图能力，支持Dispatch自动算子分发、saved_tensors_hook激活值管理和算子级注册机制，并开放自定义算子、PASS和后端能力。",{"type":17,"tag":31,"props":179,"children":180},{},[181],{"type":23,"value":182},"推理能力上，与SGLang社区合作支持MindSpore后端，适配Radix Cache等特性，升级vLLM至0.11.0并接入ACLGraph图下沉功能。科学计算方面，支持蛋白质结构预测模型Protenix，通过重计算和算子优化实现高性能训练推理。",{"type":17,"tag":77,"props":184,"children":185},{"style":79},[186],{"type":17,"tag":82,"props":187,"children":189},{"src":188,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/5.jpg",[],{"type":17,"tag":25,"props":191,"children":193},{"id":192},"社区动态",[194],{"type":23,"value":192},{"type":17,"tag":31,"props":196,"children":197},{},[198],{"type":23,"value":199},"1、昇腾社区课程发布 | MindSpore带你探索大模型",{"type":17,"tag":31,"props":201,"children":202},{},[203],{"type":23,"value":204},"1月13日，昇腾社区发布了MindSpore大模型全流程应用课程，涵盖环境搭建、预训练、微调、推理部署与调优。课程提供免费云端实验环境，支持HuggingFace、vLLM等主流生态，深入MCore架构，传授断点续训、高可用、性能调优、模型蒸馏等企业级实战技能，完成后还可获官方认证。",{"type":17,"tag":77,"props":206,"children":207},{"style":79},[208],{"type":17,"tag":82,"props":209,"children":211},{"src":210,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/6.jpg",[],{"type":17,"tag":31,"props":213,"children":214},{},[215],{"type":23,"value":216},"2、春节来OpenI复现论文，享受多卡高性能算力！",{"type":17,"tag":31,"props":218,"children":219},{},[220],{"type":23,"value":221},"2月14日至2月23日，OpenI启智社区将携手昇思MindSpore、头歌以及中国算力网、超算互联网、并行科技等算力伙伴，为高校师生和科研机构开发者开放一批高性能算力资源，支持大家安心复现论文、高效跑实验。",{"type":17,"tag":77,"props":223,"children":224},{"style":79},[225],{"type":17,"tag":82,"props":226,"children":228},{"src":227,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/7.jpg",[],{"type":17,"tag":31,"props":230,"children":231},{},[232],{"type":23,"value":233},"3、稀缺高性能算力限时开放，5大技术活动任你玩！\n昇思MindSpore联合OpenI启智社区等多家单位于2月14日至3月3日举办了“2026万码奔腾贺新春”活动。活动包含论文复现挑战、Skills炼金术、春节AI主题创作、工具链评测、年度技术复盘五大主题。本次活动旨在降低开发者技术门槛、推动国产软硬件生态创新，并进一步增强社区凝聚力。",{"type":17,"tag":77,"props":235,"children":236},{"style":79},[237],{"type":17,"tag":82,"props":238,"children":240},{"src":239,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/8.jpg",[],{"type":17,"tag":77,"props":242,"children":243},{"style":79},[244],{"type":17,"tag":82,"props":245,"children":247},{"src":246,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/91.jpg",[],{"type":17,"tag":25,"props":249,"children":251},{"id":250},"社区案例",[252],{"type":23,"value":250},{"type":17,"tag":31,"props":254,"children":255},{},[256],{"type":23,"value":257},"1、开源之夏系列活动",{"type":17,"tag":31,"props":259,"children":260},{},[261],{"type":23,"value":262},"在人才培养与社区文化建设方面，昇思MindSpore同样成果显著。在由中科院软件所发起的“开源之夏”活动中，昇思社区发布的项目共吸引103位同学提交申报，经过4个月开发期，最终22位学子成功结项。",{"type":17,"tag":31,"props":264,"children":265},{},[266],{"type":23,"value":267},"优秀成果涵盖量子计算、大模型推理优化、跨框架迁移、智能交通分析等多个前沿方向（以下为部分优秀案例，可点开链接了解详情）：",{"type":17,"tag":77,"props":269,"children":270},{"style":79},[271],{"type":17,"tag":82,"props":272,"children":274},{"src":273,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/10.jpg",[],{"type":17,"tag":31,"props":276,"children":277},{},[278],{"type":23,"value":279},"贾阔源：基于vLLM-MindSpore，探索Beam Search解码实战\n陈思源：基于MindSpore Quantum，开启量子-经典混合计算实践\n侯博森：基于昇思MindSpore的MVSNet跨框架迁移与实践\n孙文杰：基于MindSpore Quantum的对称性规约与分岔分析算法\n方泱泱：基于昇思MindSpore的YOLOv12智能交通分析实践\n黄玉含：QuCOOP量子-经典混合优化框架实践\n李振兴：基于RAG的MindSpore代码助手与智能问答系统实践",{"type":17,"tag":31,"props":281,"children":282},{},[283],{"type":23,"value":284},"2、社区“昇思同路人”系列活动",{"type":17,"tag":31,"props":286,"children":287},{},[288],{"type":23,"value":289},"与此同时，社区“昇思同路人”栏目持续记录着开发者的成长故事：",{"type":17,"tag":31,"props":291,"children":292},{},[293],{"type":23,"value":294},"第十期 讲述肖雄子彦如何搭建“学创一体”桥梁，引领学子从技术“消费者”蜕变为生态“共建者”；",{"type":17,"tag":31,"props":296,"children":297},{},[298],{"type":23,"value":299},"第十一期 讲述张威如何在校园里为昇思点燃一片“技术森林”，让技术实现“幂次方”增长；",{"type":17,"tag":31,"props":301,"children":302},{},[303],{"type":23,"value":304},"第十二期 则聚焦马欣老师的教育深耕与匠心育人，阐述AI框架如何赋能教育。",{"type":17,"tag":31,"props":306,"children":307},{},[308],{"type":23,"value":309},"3、昇思创新训练营系列活动",{"type":17,"tag":31,"props":311,"children":312},{},[313],{"type":23,"value":314},"此外，昇思创新训练营优秀案例系列分享了开发者基于MindSpore微调大模型、构建智能旅游助手的全流程实践，项目代码已开源，欢迎开发者体验！",{"type":17,"tag":77,"props":316,"children":317},{"style":79},[318],{"type":17,"tag":82,"props":319,"children":321},{"src":320,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/11.jpg",[],{"type":17,"tag":31,"props":323,"children":324},{},[325],{"type":23,"value":326},"这些鲜活的案例共同勾勒出昇思MindSpore社区以技术育人为核、以同路人为荣的浓厚氛围。",{"type":17,"tag":25,"props":328,"children":330},{"id":329},"技术分享",[331],{"type":23,"value":329},{"type":17,"tag":333,"props":334,"children":336},"h3",{"id":335},"_1-大模型训练hyperparallel架构发布",[337],{"type":23,"value":338},"1 大模型训练：HyperParallel架构发布",{"type":17,"tag":31,"props":340,"children":341},{},[342],{"type":23,"value":343},"本季度正式推出为超节点设计的HyperParallel架构，包含三大核心技术：",{"type":17,"tag":31,"props":345,"children":346},{},[347],{"type":23,"value":348},"HyperShard声明式并行：用户只需声明输入、输出和参数的切分方式，框架自动完成分布式策略推导与资源调度，实现“编写即单卡，运行即分布式”的体验。",{"type":17,"tag":31,"props":350,"children":351},{},[352],{"type":23,"value":353},"HyperOffload多级智能卸载：在编译期静态分析计算图，将长生命周期张量自动卸载至主机内存，降低峰值显存占用，支持更长序列训练。经测试，模型训练吞吐平均提升超过10%。",{"type":17,"tag":77,"props":355,"children":356},{"style":79},[357],{"type":17,"tag":82,"props":358,"children":360},{"src":359,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/12.jpg",[],{"type":17,"tag":31,"props":362,"children":363},{},[364],{"type":23,"value":365},"HyperMPMD非规则异构并行：应对模型结构不规则与硬件异构性挑战，实现灵活高效的资源调度。",{"type":17,"tag":333,"props":367,"children":369},{"id":368},"_2-大模型推理sglang后端集成与vllm升级",[370],{"type":23,"value":371},"2 大模型推理：SGLang后端集成与vLLM升级",{"type":17,"tag":31,"props":373,"children":374},{},[375],{"type":23,"value":376},"与SGLang社区深度合作，正式合入MindSpore推理后端代码，该方案通过封装MindSpore模型复用SGLang原生组件，借助DLPack实现张量零拷贝转换，打通双框架协同，支持数据并行、PD分离、投机解码等特性。",{"type":17,"tag":31,"props":378,"children":379},{},[380],{"type":23,"value":381},"升级适配vLLM v0.11.0，接入昇腾ACLGraph图下沉功能。实测DeepSeek-V3/R1 W4A8量化推理服务中，整网吞吐提升约5%，算子下发时延由30ms降至10ms。",{"type":17,"tag":31,"props":383,"children":384},{},[385],{"type":23,"value":386},"通过DLPack实现MindSpore与PyTorch张量零拷贝转换，跨框架协同效率大幅提升。",{"type":17,"tag":77,"props":388,"children":389},{"style":79},[390],{"type":17,"tag":82,"props":391,"children":393},{"src":392,"style":85,"alt":7},"/category/information/news/banner/2026-4-9/13.jpg",[],{"type":17,"tag":333,"props":395,"children":397},{"id":396},"_3-基础框架演进动态图增强与自定义能力开放",[398],{"type":23,"value":399},"3 基础框架演进：动态图增强与自定义能力开放",{"type":17,"tag":31,"props":401,"children":402},{},[403],{"type":23,"value":404},"动态图能力增强：新增Dispatch自动算子分发、saved_tensors_hook激活值管理、算子级注册机制，提升多设备协同易用性与显存优化灵活性。",{"type":17,"tag":31,"props":406,"children":407},{},[408],{"type":23,"value":409},"自定义能力全面开放：提供自定义算子（极简C++接口，自动享受内存复用与异步调度）、自定义PASS（开放图优化接口，支持算子融合等变换）、自定义后端（标准化接口快速适配新型硬件），从“封闭工具箱”向“开放创新平台”演进。",{"type":17,"tag":333,"props":411,"children":413},{"id":412},"_4-科学计算套件蛋白质结构预测模型protenix深度优化",[414],{"type":23,"value":415},"4 科学计算套件：蛋白质结构预测模型Protenix深度优化",{"type":17,"tag":31,"props":417,"children":418},{},[419],{"type":23,"value":420},"支持AlphaFold3的高性能开源复现模型Protenix v0.5，在昇腾硬件上完成训推全流程适配。",{"type":17,"tag":31,"props":422,"children":423},{},[424],{"type":23,"value":425},"训练方面：通过重计算优化，最大支持序列长度从64提升至768，动态显存峰值从20152MB降至7025MB。",{"type":17,"tag":31,"props":427,"children":428},{},[429],{"type":23,"value":430},"推理方面：融合算子优化后单卡推理性能提升超100%；长序列推理通过分块计算将单卡推理长度极限提升至3000以上；端到端加速比达57%。",{"type":17,"tag":333,"props":432,"children":434},{"id":433},"_5-开发者工具与实战指南",[435],{"type":23,"value":436},"5 开发者工具与实战指南",{"type":17,"tag":31,"props":438,"children":439},{},[440],{"type":23,"value":441},"数据流水线与混合精度优化：针对昇腾AI处理器，总结“三板斧”——数据流水线并行（num_parallel_workers）、自动混合精度（amp_level=\"O2\"）、数据下沉（dataset_sink_mode=True），有效解决IO瓶颈，提升训练速度。",{"type":17,"tag":31,"props":443,"children":444},{},[445],{"type":23,"value":446},"半自动并行工具Shard：深度解读shard接口的三阶段机制（策略锚定→BFS传播→冲突重排），帮助用户在自动并行与手动调优间取得平衡。",{"type":17,"tag":31,"props":448,"children":449},{},[450],{"type":23,"value":451},"优化器架构：MindSpore优化器通过静态图编译与整图下沉将参数更新闭环在设备侧，支持融合算子、解耦权重衰减、ZeRO优化器并行等特性，实现从单卡到千亿级模型的极致性能。",{"type":17,"tag":31,"props":453,"children":454},{},[455],{"type":23,"value":456},"自动并行实战：一键配置并行模式（数据并行/半自动/全自动），无需修改网络结构，让开发者聚焦模型创新。",{"type":17,"tag":31,"props":458,"children":459},{},[460],{"type":23,"value":461},"开放麦：提供零基础AI艺术生成器教程，基于VAE构建完整代码，涵盖训练、保存、生成与可视化流程。",{"type":17,"tag":31,"props":463,"children":464},{},[465,467],{"type":23,"value":466},"💡 更多技术解读，欢迎访问昇思社区官网：",{"type":17,"tag":468,"props":469,"children":473},"a",{"href":470,"rel":471},"https://discuss.mindspore.cn/",[472],"nofollow",[474],{"type":23,"value":470},{"type":17,"tag":25,"props":476,"children":478},{"id":477},"感谢每一位朋友开发者的支持",[479],{"type":23,"value":480},"感谢每一位朋友、开发者的支持",{"type":17,"tag":31,"props":482,"children":483},{},[484],{"type":23,"value":485},"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":31,"props":487,"children":488},{},[489],{"type":23,"value":490},"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":492,"depth":492,"links":493},4,[494,496,497,498,499,500,508],{"id":27,"depth":495,"text":27},2,{"id":58,"depth":495,"text":58},{"id":109,"depth":495,"text":109},{"id":192,"depth":495,"text":192},{"id":250,"depth":495,"text":250},{"id":329,"depth":495,"text":329,"children":501},[502,504,505,506,507],{"id":335,"depth":503,"text":338},3,{"id":368,"depth":503,"text":371},{"id":396,"depth":503,"text":399},{"id":412,"depth":503,"text":415},{"id":433,"depth":503,"text":436},{"id":477,"depth":495,"text":480},"markdown","content:news:zh:2026-4-9.md","content","news/zh/2026-4-9.md","news/zh/2026-4-9","md",1776506061322]