[{"data":1,"prerenderedAt":274},["ShallowReactive",2],{"content-query-aYi5CNlRqq":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":268,"_id":269,"_source":270,"_file":271,"_stem":272,"_extension":273},"/version-updates/zh/453","zh",false,"","MindSpore 1.2：国内首个支持千亿参数大模型训练的AI计算框架","AI计算框架MindSpore 1.2正式发布","2021-04-29","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/07/79ea5b0bb9064e7fb245bb641e480e30.png","version-updates",{"type":14,"children":15,"toc":265},"root",[16,24,30,38,52,62,67,74,79,84,91,98,103,108,115,126,136,141,150,157,162,172,179,184,191,196,206,213,218,227,232,243,248,253,258],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"mindspore-12国内首个支持千亿参数大模型训练的ai计算框架",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":29},"国内首个支持千亿参数大模型训练的AI计算框架MindSpore 1.2正式发布，无论是动态图分布式训练效率的大幅提升，还是一键模型迁移、CA鲁棒性标准达标、深度分子模拟及量子机器学习等，都能让AI开发者尽享AI开发。",{"type":17,"tag":25,"props":31,"children":32},{},[33],{"type":17,"tag":34,"props":35,"children":37},"img",{"alt":7,"src":36},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/085593fda5f84195b19348485a2b4001.jpg",[],{"type":17,"tag":25,"props":39,"children":40},{},[41,50],{"type":17,"tag":42,"props":43,"children":47},"a",{"href":44,"rel":45},"https://mp.weixin.qq.com/s/MQO3EWoWOhOKZDq7m1Bk7w",[46],"nofollow",[48],{"type":23,"value":49},"老规矩上视频！",{"type":23,"value":51},"（请点击“老规矩上视频！”并在微信里进行视频观看）",{"type":17,"tag":25,"props":53,"children":54},{},[55],{"type":17,"tag":42,"props":56,"children":59},{"href":57,"rel":58},"https://mp.weixin.qq.com/s?__biz=MzAxMDA1MDM0NQ==&mid=2247529356&idx=1&sn=05f66800b55ee3bc5bd90cf07605967a&chksm=9b547dcaac23f4dc8f90c2e1729716634dd0f43e2f52c26d227ff5ec02853dc6992df0df43b6&scene=21#wechat_redirect",[46],[60],{"type":23,"value":61},"MindSpore开源周年狂欢，量子机器学习与深度分子模拟等巨量新特性来袭！",{"type":17,"tag":25,"props":63,"children":64},{},[65],{"type":23,"value":66},"先回顾下我们在一周年中发布的新特性，再来看看这次新版本还有哪些新特性~",{"type":17,"tag":25,"props":68,"children":69},{},[70],{"type":17,"tag":34,"props":71,"children":73},{"alt":7,"src":72},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/b4d6640c23a44f3e88848d73c757368e.jpg",[],{"type":17,"tag":25,"props":75,"children":76},{},[77],{"type":23,"value":78},"如何在大集群下高效训练千亿至万亿参数模型？当前主要面临两个挑战：切分难——用户需要综合考虑参数量、计算量、计算类型、集群带宽拓扑和样本数量等，才能设计出性能较优的并行切分策略；编码难——模型编码除了要考虑算法以外，还需要编写大量并行切分和通信代码。",{"type":17,"tag":25,"props":80,"children":81},{},[82],{"type":23,"value":83},"MindSpore采用多维度自动并行，通过数据并行、模型并行、Pipeline并行、异构并行、重复计算、高效内存复用及拓扑感知调度，降低通信开销，实现整体迭代时间最小（计算时间+通信时间）。",{"type":17,"tag":25,"props":85,"children":86},{},[87],{"type":17,"tag":34,"props":88,"children":90},{"alt":7,"src":89},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/78182b804f744eeabeea9cd80119e2f4.jpg",[],{"type":17,"tag":25,"props":92,"children":93},{},[94],{"type":17,"tag":34,"props":95,"children":97},{"alt":7,"src":96},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/72e0ca01f659417094e5e45295b00f6b.jpg",[],{"type":17,"tag":25,"props":99,"children":100},{},[101],{"type":23,"value":102},"相较于移动终端设备，IoT设备上系统资源有限，对ROM空间占用、运行时内存和功耗要求较高，为了使得IoT设备上也能方便快速的部署AI应用，MindSpore Lite此次正式开源了代码自动生成工具Codegen，旨在将模型编译为极简的代码，用于推理。",{"type":17,"tag":25,"props":104,"children":105},{},[106],{"type":23,"value":107},"Codegen可对接MindSpore Lite的NNACL算子库和ARM的CMSIS算子库，支持生成可在X86/ARM64/ARM32A/ARM32M平台部署的推理代码。",{"type":17,"tag":25,"props":109,"children":110},{},[111],{"type":17,"tag":34,"props":112,"children":114},{"alt":7,"src":113},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/a2b12ba721b548c9962da8f61139f717.jpg",[],{"type":17,"tag":25,"props":116,"children":117},{},[118,124],{"type":17,"tag":119,"props":120,"children":121},"strong",{},[122],{"type":23,"value":123},"易上手",{"type":23,"value":125},"：用户输入ms模型文件，并指定推理执行硬件平台，即可生成与该模型执行相关的代码、编译构建脚本。",{"type":17,"tag":25,"props":127,"children":128},{},[129,134],{"type":17,"tag":119,"props":130,"children":131},{},[132],{"type":23,"value":133},"极致优化",{"type":23,"value":135},"：无图引擎，模型即代码，极小ROM占用完成推理。算子数据重排、内存分配移至离线完成，节省了初始化时间；同时引入Conv & BN & Pooling融合，Conv分块计算等技术，达到极致性能。我们的方案在华为内部应用广泛，比如在Huawei Watch GT系列的智慧抬腕亮屏算法中，推理内存占用仅1KB，时延低至5ms。",{"type":17,"tag":25,"props":137,"children":138},{},[139],{"type":23,"value":140},"官网快速入门：",{"type":17,"tag":25,"props":142,"children":143},{},[144],{"type":17,"tag":42,"props":145,"children":148},{"href":146,"rel":147},"https://www.mindspore.cn/",[46],[149],{"type":23,"value":146},{"type":17,"tag":25,"props":151,"children":152},{},[153],{"type":17,"tag":34,"props":154,"children":156},{"alt":7,"src":155},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/a2a2ac5f40ee4886ac3c5fd6cca897a9.jpg",[],{"type":17,"tag":25,"props":158,"children":159},{},[160],{"type":23,"value":161},"MindSpore在MindInsight部件中集成了的可解释AI能力：显著图可视化、反事实解释和可解释AI评估体系（度量体系），旨在帮忙开发者更好进行模型调优。",{"type":17,"tag":25,"props":163,"children":164},{},[165,170],{"type":17,"tag":119,"props":166,"children":167},{},[168],{"type":23,"value":169},"显著图可视化和反事实解释",{"type":23,"value":171},"：能够为图片分类模型的推理结果，标识影响分类结果的关键特征，方便理解分类依据，帮助分析定位分类错误原因，加速模型调优。",{"type":17,"tag":25,"props":173,"children":174},{},[175],{"type":17,"tag":34,"props":176,"children":178},{"alt":7,"src":177},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/faa84cbfd06944deafc0b0113838d931.jpg",[],{"type":17,"tag":25,"props":180,"children":181},{},[182],{"type":23,"value":183},"这个例子中，有个错误预测标签“person”，根据显著图可视化结果，高亮区域在摩托车的前部，便于针对性的分析误判的可能原因。",{"type":17,"tag":25,"props":185,"children":186},{},[187],{"type":17,"tag":34,"props":188,"children":190},{"alt":7,"src":189},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/a295ede484de416cb3db8420bbbb753b.jpg",[],{"type":17,"tag":25,"props":192,"children":193},{},[194],{"type":23,"value":195},"在这个例子中，图片被错分为“cat”，使用基于层级遮掩的反事实解释，发现对这个分类结果影响最可能的区域是右边遮掩后留下的区域，便于用户发现判断错误的缘由，从而帮助调优。",{"type":17,"tag":25,"props":197,"children":198},{},[199,204],{"type":17,"tag":119,"props":200,"children":201},{},[202],{"type":23,"value":203},"评估体系",{"type":23,"value":205},"：面向各种解释方法，给出在具体场景下，选中解释方法在不同度量方法下的解释效果得分，助力用户选择最优的解释方法，从而更好的帮助模型调优。",{"type":17,"tag":25,"props":207,"children":208},{},[209],{"type":17,"tag":34,"props":210,"children":212},{"alt":7,"src":211},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/552ce22162f44cc799bcb025f2f1e07a.jpg",[],{"type":17,"tag":25,"props":214,"children":215},{},[216],{"type":23,"value":217},"查看教程：",{"type":17,"tag":25,"props":219,"children":220},{},[221],{"type":17,"tag":42,"props":222,"children":225},{"href":223,"rel":224},"https://www.mindspore.cn/tutorial/training/zh-CN/r1.1/advanced_use/model_explaination.html",[46],[226],{"type":23,"value":223},{"type":17,"tag":25,"props":228,"children":229},{},[230],{"type":23,"value":231},"MindSpore官方资料",{"type":17,"tag":25,"props":233,"children":234},{},[235,237],{"type":23,"value":236},"GitHub : ",{"type":17,"tag":42,"props":238,"children":241},{"href":239,"rel":240},"https://github.com/mindspore-ai/mindspore",[46],[242],{"type":23,"value":239},{"type":17,"tag":25,"props":244,"children":245},{},[246],{"type":23,"value":247},"Gitee:[https : //gitee.com/mindspore/mindspore](https : //gitee.com/mindspore/mindspore)",{"type":17,"tag":25,"props":249,"children":250},{},[251],{"type":23,"value":252},"官方QQ群 : 871543426",{"type":17,"tag":25,"props":254,"children":255},{},[256],{"type":23,"value":257},"扫描下方二维码加入MindSpore项目↓",{"type":17,"tag":25,"props":259,"children":260},{},[261],{"type":17,"tag":34,"props":262,"children":264},{"alt":7,"src":263},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/04/29/ef29f5fc0b5f4dbc8c007674043d0f5b.jpg",[],{"title":7,"searchDepth":266,"depth":266,"links":267},4,[],"markdown","content:version-updates:zh:453.md","content","version-updates/zh/453.md","version-updates/zh/453","md",1776506145734]