[{"data":1,"prerenderedAt":232},["ShallowReactive",2],{"content-query-UN2D2sboer":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},"/news/zh/3758","zh",false,"","独家秘籍：探索昇思MindSpore如何让SOTA模型迁得快、对得齐","昇思MindSpore支持大模型训练Day0迁移、推理一键部署，携手开发者共筑开源生态。","2025-06-12","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/06/13/9d5dd3e4a39d44bb87d0266aad0580ef.png","news",{"type":14,"children":15,"toc":220},"root",[16,24,30,35,40,48,53,58,67,77,82,87,92,112,120,129,134,139,144,149,157,166,171,176,181,199,204,209],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"独家秘籍探索昇思mindspore如何让sota模型迁得快对得齐",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":29},"大模型发展日新月异，新的大模型层出不穷，参数规模从十亿级跃升至万亿级，模型结构从稠密到稀疏，应用范式走向多模态、Agent… 如此多的模型和技术如何在昇腾上快速的使能和体验，是昇腾开发者的核心诉求。昇思MindSpore “海纳百川”，全面融入主流生态，全面优化训练与推理全流程开发体验，让开发者用极小成本完成SOTA大模型的迁移，昇思MindSpore需攻克两大关键挑战：",{"type":17,"tag":25,"props":31,"children":32},{},[33],{"type":23,"value":34},"1. 迁得快：让三方框架模型\"零成本\"迁移，避免重复造轮子，同时模型精度完全对齐。",{"type":17,"tag":25,"props":36,"children":37},{},[38],{"type":23,"value":39},"2. 部署快：训转推全流程自动化，让大模型部署像执行一行命令一般敏捷高效。",{"type":17,"tag":25,"props":41,"children":42},{},[43],{"type":17,"tag":44,"props":45,"children":47},"img",{"alt":7,"src":46},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/06/13/fffe427f9cfb47db9e6a69417e3b242b.png",[],{"type":17,"tag":25,"props":49,"children":50},{},[51],{"type":23,"value":52},"Figure 1 MindSpore生态快速迁移解决方案的技术架构",{"type":17,"tag":25,"props":54,"children":55},{},[56],{"type":23,"value":57},"接下来，我们将揭开昇思MindSpore的破局之道。",{"type":17,"tag":25,"props":59,"children":60},{},[61],{"type":17,"tag":62,"props":63,"children":64},"strong",{},[65],{"type":23,"value":66},"01",{"type":17,"tag":68,"props":69,"children":71},"h4",{"id":70},"支持训练day0迁移构建跨框架的无感智能翻译能力",[72],{"type":17,"tag":62,"props":73,"children":74},{},[75],{"type":23,"value":76},"支持训练Day0迁移，构建跨框架的\"无感智能翻译\"能力",{"type":17,"tag":25,"props":78,"children":79},{},[80],{"type":23,"value":81},"当大模型架构日新月异，开发者最怕被生态绑定。昇思MindSpore通过三重兼容术打通主流技术栈，支持主流加速库模型0代码迁移，通过精度自动对比工具实现跨框架、跨版本、跨策略快速调优，精度对齐原模型，实现在分布式并行策略保持不变的情况下，训练性能提升5%+。",{"type":17,"tag":25,"props":83,"children":84},{},[85],{"type":23,"value":86},"在训练生态方面，通过MindSpeed/Megatron桥接层实现PyTorch模型零代码迁移，训练脚本可直接运行；通过动态图能力重构，昇思让PyTorch开发者获得\"原生体验\"，同时借力MSAdapter工具自动转换95%以上接口，主流模型如DeepSeek、Pangu等迁移损耗逼近于零。",{"type":17,"tag":25,"props":88,"children":89},{},[90],{"type":23,"value":91},"迁移只是起点，性能才是硬道理。昇思MindSpore构建了框架差异化技术，进一步提升模型训练效率：",{"type":17,"tag":93,"props":94,"children":95},"ul",{},[96,102,107],{"type":17,"tag":97,"props":98,"children":99},"li",{},[100],{"type":23,"value":101},"动态图编译优化：多级流水线+即时编译（JIT），单卡训练效率提升40%",{"type":17,"tag":97,"props":103,"children":104},{},[105],{"type":23,"value":106},"分布式智能调优：自动负载均衡工具解决万卡训练\"木桶效应\"，线性度突破96%",{"type":17,"tag":97,"props":108,"children":109},{},[110],{"type":23,"value":111},"逐层精度对齐：在昇腾硬件上实现主流模型逐层0误差，告别玄学调参",{"type":17,"tag":25,"props":113,"children":114},{},[115],{"type":17,"tag":62,"props":116,"children":117},{},[118],{"type":23,"value":119},"02",{"type":17,"tag":68,"props":121,"children":123},{"id":122},"支持推理一键部署分钟级拉起模型服务",[124],{"type":17,"tag":62,"props":125,"children":126},{},[127],{"type":23,"value":128},"支持推理一键部署，分钟级拉起模型服务",{"type":17,"tag":25,"props":130,"children":131},{},[132],{"type":23,"value":133},"在推理部署方面，借助vLLM-MindSpore插件实现HuggingFace模型半小时部署上线。面对千亿参数模型，昇思用三层部署范式重构推理链路：",{"type":17,"tag":25,"props":135,"children":136},{},[137],{"type":23,"value":138},"1. HF权重兼容：支持HuggingFace权重直接加载，无需格式转换；通过vLLM-MindSpore插件实现分钟级服务化",{"type":17,"tag":25,"props":140,"children":141},{},[142],{"type":23,"value":143},"2. 模型开箱即用：支持业界主流模型开箱即用，DeepSeek、Pangu、Qwen等20+模型已上线",{"type":17,"tag":25,"props":145,"children":146},{},[147],{"type":23,"value":148},"3. 启动时延优化：权重加载耗时降低80%（百亿模型小于30秒）；图编译时延压缩至毫秒级",{"type":17,"tag":25,"props":150,"children":151},{},[152],{"type":17,"tag":62,"props":153,"children":154},{},[155],{"type":23,"value":156},"03",{"type":17,"tag":68,"props":158,"children":160},{"id":159},"开源开放与社区开发者共建共赢",[161],{"type":17,"tag":62,"props":162,"children":163},{},[164],{"type":23,"value":165},"开源开放：与社区开发者共建共赢",{"type":17,"tag":25,"props":167,"children":168},{},[169],{"type":23,"value":170},"在人工智能的沃土之上，昇思MindSpore已生长为一片葱郁的开发者雨林。",{"type":17,"tag":25,"props":172,"children":173},{},[174],{"type":23,"value":175},"自2020年3月28日播下开源的种子，这片技术雨林已孕育出50多个主流大模型。AI框架作为基础软件，以其根系持续释放创新养分。昇思MindSpore累计获得1200万次下载，开发者枝芽遍布130个国家和地区土壤，在2400座城市吐露生机，超过四万六千名开发者如蜂群般共舞其间，11万+行代码代码合入请求浇灌创新土壤，1700多篇学术硕果垂坠于智慧枝头。",{"type":17,"tag":25,"props":177,"children":178},{},[179],{"type":23,"value":180},"面向开发者，昇思的终极武器是开源开放的创新生态：",{"type":17,"tag":93,"props":182,"children":183},{},[184,189,194],{"type":17,"tag":97,"props":185,"children":186},{},[187],{"type":23,"value":188},"社区共治：理事会+SIG组双轮驱动，企业/高校共定技术路线",{"type":17,"tag":97,"props":190,"children":191},{},[192],{"type":23,"value":193},"资源普惠：提供免费算力平台（MindSpore大模型平台）",{"type":17,"tag":97,"props":195,"children":196},{},[197],{"type":23,"value":198},"知识共享：20+技术SIG组覆盖AI前沿领域，开发者可参与核心开发",{"type":17,"tag":25,"props":200,"children":201},{},[202],{"type":23,"value":203},"未来的AI软件基座，必属于开源开放与极简高效共舞的时代。",{"type":17,"tag":25,"props":205,"children":206},{},[207],{"type":23,"value":208},"文章转载雷锋网：",{"type":17,"tag":25,"props":210,"children":211},{},[212],{"type":17,"tag":213,"props":214,"children":218},"a",{"href":215,"rel":216},"https://mp.weixin.qq.com/s?__biz=MTM2ODM0ODYyMQ==&mid=2651733647&idx=2&sn=cd05bf9ba09bdc7f426167ac65cae646&scene=21#wechat_redirect",[217],"nofollow",[219],{"type":23,"value":8},{"title":7,"searchDepth":221,"depth":221,"links":222},4,[223,224,225],{"id":70,"depth":221,"text":76},{"id":122,"depth":221,"text":128},{"id":159,"depth":221,"text":165},"markdown","content:news:zh:3758.md","content","news/zh/3758.md","news/zh/3758","md",1776506088826]