[{"data":1,"prerenderedAt":82},["ShallowReactive",2],{"content-query-okBf9K93Pe":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":76,"_id":77,"_source":78,"_file":79,"_stem":80,"_extension":81},"/news/zh/3842","zh",false,"","OpenBMB 开源社区 BMTrain携手昇思MindSpore，加速大模型高效训练","为开发者提供高性能、高易用的大模型训练解决方案","2025-09-22","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/09/29/a0917d35cbea4e68a356042c767086b0.png","news",{"type":14,"children":15,"toc":73},"root",[16,24,30,43,48,53,58,63,68],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"openbmb-开源社区-bmtrain携手昇思mindspore加速大模型高效训练",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":29},"近日，全场景AI框架昇思MindSpore与OpenBMB 开源社区推出的高效大模型训练引擎BMTrain进一步达成合作，双方将持续共同推动大模型预训练与微调技术的创新与落地。此次合作旨在将昇思MindSpore AI框架的全场景协同、多维混合分布式并行能力等和BMTrain的高效训练特性结合，为开发者提供高性能、高易用的大模型训练解决方案。",{"type":17,"tag":25,"props":31,"children":32},{},[33,35],{"type":23,"value":34},"代码链接：",{"type":17,"tag":36,"props":37,"children":41},"a",{"href":38,"rel":39},"https://github.com/OpenBMB/BMTrain/tree/mindspore#",[40],"nofollow",[42],{"type":23,"value":38},{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":23,"value":47},"根据双方合作显示，昇思MindSpore与BMTrain将聚焦于大模型训练技术的深度融合。BMTrain作为一款专为大规模模型训练设计的轻量级引擎，其高效的内存管理和分布式优化能力与MindSpore的自动并行、动静统一等特性形成互补。双方将通过接口兼容、联合优化等方式，实现训练成本的大幅降低和开发效率的显著提升。",{"type":17,"tag":25,"props":49,"children":50},{},[51],{"type":23,"value":52},"合作后，开发者可基于MindSpore框架直接调用BMTrain的优化能力，实现训练成本大幅降低，同时支持万亿级参数模型的高效训练。此举将加速大模型在科研与产业界的应用，以及长序列处理、多模态理解等复杂模型场景中发挥关键作用。",{"type":17,"tag":25,"props":54,"children":55},{},[56],{"type":23,"value":57},"昇思MindSpore是支持全自动并行的AI框架，具备动静统一、训推一体核心特性。其支持图算融合、自动微分、多维混合并行等技术，显著提升训练性能。目前，MindSpore已兼容业界主要生态接口，支持LLaMA、Qwen、DeepSeek等超百个主流大模型，并在昇腾AI硬件上实现性能领先。",{"type":17,"tag":25,"props":59,"children":60},{},[61],{"type":23,"value":62},"BMTrain则是一款专注于大模型预训练与微调的轻量级训练引擎，通过动态内存优化、通信压缩和梯度策略创新，大幅降低计算与存储开销。其相比DeepSpeed等框架可节省90%训练成本，同时支持超长序列训练和高效异构计算调度。",{"type":17,"tag":25,"props":64,"children":65},{},[66],{"type":23,"value":67},"未来，双方计划在长序列训练优化、科学智能融合、生态共建等领域深化合作。具体包括结合MindSpore的上下文并行与BMTrain的内存管理，突破万亿Token序列训练瓶颈；拓展AI+科学计算场景，支持生物医药、气象预测等领域的超大模型训练；推动BMTrain与MindSpore社区共享工具链，降低开发者迁移与调优门槛。",{"type":17,"tag":25,"props":69,"children":70},{},[71],{"type":23,"value":72},"此次合作标志着自主创新的AI框架与训练引擎的协同创新进入新阶段，将为全球开发者提供更高效、更经济的大模型基础设施。",{"title":7,"searchDepth":74,"depth":74,"links":75},4,[],"markdown","content:news:zh:3842.md","content","news/zh/3842.md","news/zh/3842","md",1776506090860]