[{"data":1,"prerenderedAt":599},["ShallowReactive",2],{"content-query-MJFn3iphJ7":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":593,"_id":594,"_source":595,"_file":596,"_stem":597,"_extension":598},"/version-updates/zh/653","zh",false,"","新版本 | MindSpore 1.3，打造无所不在的智能，诠释可以信赖的开源","MindSpore 1.3最新版本发布，全场景多项能力提升！","2021-07-16","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/9604f707d9124f6eab130e4d492ceab1.png","version-updates",{"type":14,"children":15,"toc":590},"root",[16,24,33,38,56,61,95,102,119,124,134,142,150,187,192,201,209,214,221,226,233,243,248,253,260,265,272,277,284,288,298,306,311,323,330,335,345,353,358,373,381,386,391,398,403,410,418,423,431,436,441,451,461,469,487,494,502,507,518,525,532,540,552,557,566,571,578,583],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"新版本-mindspore-13打造无所不在的智能诠释可以信赖的开源",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":17,"tag":29,"props":30,"children":32},"img",{"alt":7,"src":31},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/2a08fcc38e0d436d9efee1af592af8ed.gif",[],{"type":17,"tag":25,"props":34,"children":35},{},[36],{"type":23,"value":37},"夏日的阳光已经洒在树梢枝畔，奋斗了又3个月的MindSpore社区开发者们，带着无比的兴奋，为大家献上全新的1.3版本！日益精进的我们，在这个版本中为大家带来了全新的MindSpore Federated能力，解锁了支撑盘古千亿稠密大模型的众多关键特性、以及面向更多类型硬件的推理优化、图算融合、简易部署等新工艺，犹如开源时所誓言，持续不断的为AI开发者带来惊喜。下面就带大家快速浏览1.3版本的关键特性。",{"type":17,"tag":25,"props":39,"children":40},{},[41,45,47],{"type":17,"tag":29,"props":42,"children":44},{"alt":7,"src":43},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/5aa90201c12e40858bf011d9ccebdc77.png",[],{"type":23,"value":46}," ",{"type":17,"tag":48,"props":49,"children":53},"a",{"href":50,"rel":51},"https://mp.weixin.qq.com/s/9N_Ib8ZbgbVVEn-7R8zpKg",[52],"nofollow",[54],{"type":23,"value":55},"请在微信内进行观看",{"type":17,"tag":25,"props":57,"children":58},{},[59],{"type":23,"value":60},"联邦学习是一种加密的分布式机器学习技术，它是指参与联邦学习的各用户在不共享本地数据的前提下共建AI模型，主要用于解决数据隐私和数据孤岛的问题。MindSpore Federated优先专注于大规模参与方的横向联邦的应用场景。",{"type":17,"tag":25,"props":62,"children":63},{},[64,66,72,74,79,81,86,88,93],{"type":23,"value":65},"端云场景下的参与方是非常大量的手机或者IoT设备，其分布规模和设备不可靠性带来了系统异构、通信效率、隐私安全和标签缺失等挑战。MindSpore Federated设计了松耦合分布式服务器、限时通信模块、容忍退出的安全聚合等模块，使得",{"type":17,"tag":67,"props":68,"children":69},"strong",{},[70],{"type":23,"value":71},"任何时候",{"type":23,"value":73},"只要部分客户端可用，都可以",{"type":17,"tag":67,"props":75,"children":76},{},[77],{"type":23,"value":78},"进行联邦学习任务",{"type":23,"value":80},"，并",{"type":17,"tag":67,"props":82,"children":83},{},[84],{"type":23,"value":85},"解决了系统异构带来的“长尾效应”",{"type":23,"value":87},"，",{"type":17,"tag":67,"props":89,"children":90},{},[91],{"type":23,"value":92},"提高了学习效率",{"type":23,"value":94},"。",{"type":17,"tag":25,"props":96,"children":97},{},[98],{"type":17,"tag":29,"props":99,"children":101},{"alt":7,"src":100},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/8fe3e4595c14480b8fec62af8047fd60.jpg",[],{"type":17,"tag":25,"props":103,"children":104},{},[105,107,112,113,118],{"type":23,"value":106},"搭载于MindSpore端边云全场景统一的基础架构，MindSpore Federated将会作为华为终端的核心竞争力之一，",{"type":17,"tag":67,"props":108,"children":109},{},[110],{"type":23,"value":111},"在保护用户隐私的前提下",{"type":23,"value":87},{"type":17,"tag":67,"props":114,"children":115},{},[116],{"type":23,"value":117},"提供更具有个性化的用户信息决策",{"type":23,"value":94},{"type":17,"tag":25,"props":120,"children":121},{},[122],{"type":23,"value":123},"查看介绍：",{"type":17,"tag":25,"props":125,"children":126},{},[127],{"type":17,"tag":48,"props":128,"children":131},{"href":129,"rel":130},"https://gitee.com/mindspore/docs/blob/r1.3/docs/federated/summarize%5C_federated.md",[52],[132],{"type":23,"value":133},"https://gitee.com/mindspore/docs/blob/r1.3/docs/federated/summarize\\_federated.md",{"type":17,"tag":25,"props":135,"children":136},{},[137],{"type":17,"tag":67,"props":138,"children":139},{},[140],{"type":23,"value":141},"盘古千亿稠密大模型关键特性开源",{"type":17,"tag":25,"props":143,"children":144},{},[145],{"type":17,"tag":67,"props":146,"children":147},{},[148],{"type":23,"value":149},"0. 中文语言理解任务评测全球第一",{"type":17,"tag":25,"props":151,"children":152},{},[153,158,160,165,166,171,173,178,180,185],{"type":17,"tag":67,"props":154,"children":155},{},[156],{"type":23,"value":157},"全球最大中文语言预训练模型“鹏程.盘古”",{"type":23,"value":159},"，在2021年4月23日权威的中文语言理解评测基准CLUE榜单中，总成绩及阅读理解、分类任务单项均",{"type":17,"tag":67,"props":161,"children":162},{},[163],{"type":23,"value":164},"排名第一",{"type":23,"value":87},{"type":17,"tag":67,"props":167,"children":168},{},[169],{"type":23,"value":170},"刷新",{"type":23,"value":172},"三项榜单",{"type":17,"tag":67,"props":174,"children":175},{},[176],{"type":23,"value":177},"纪录",{"type":23,"value":179},"，总成绩得分83；在NLPCC2018文本摘要任务中，取得了Rouge平均分0.53的",{"type":17,"tag":67,"props":181,"children":182},{},[183],{"type":23,"value":184},"业界最佳成绩",{"type":23,"value":186},"，超越第二名百分之六十。",{"type":17,"tag":25,"props":188,"children":189},{},[190],{"type":23,"value":191},"模型详细介绍请参考MindSpore首席架构师金雪锋的分析文章：",{"type":17,"tag":25,"props":193,"children":194},{},[195],{"type":17,"tag":48,"props":196,"children":199},{"href":197,"rel":198},"https://www.zhihu.com/question/456443707/answer/1856014437",[52],[200],{"type":23,"value":197},{"type":17,"tag":25,"props":202,"children":203},{},[204],{"type":17,"tag":67,"props":205,"children":206},{},[207],{"type":23,"value":208},"1. 超大模型关键特性之——分布式推理及在线部署",{"type":17,"tag":25,"props":210,"children":211},{},[212],{"type":23,"value":213},"当大模型推理参数量较大难以部署到一张卡上时，MindSpore可通过OP-Level和PipeLine模型并行将大模型切分到多卡。我们实现了鹏程.盘古2000亿级参数量的模型部署。",{"type":17,"tag":25,"props":215,"children":216},{},[217],{"type":17,"tag":29,"props":218,"children":220},{"alt":7,"src":219},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/9b9b7bbbeb184947b654bc5c175772e3.jpg",[],{"type":17,"tag":25,"props":222,"children":223},{},[224],{"type":23,"value":225},"MindSpore Serving支持分布式模型的在线推理部署，根据模型并行策略切分为多个小图，每个卡部署两个图的各一个小图。下图中，每个卡通过一个Agent进程管理图加载和执行，卡之间通过HCCL进行高速通信，分布式主worker进程和各个Agent进程之间通过gRPC通信，用于请求数据的发送和请求结果的接受。",{"type":17,"tag":25,"props":227,"children":228},{},[229],{"type":17,"tag":29,"props":230,"children":232},{"alt":7,"src":231},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/56f73a0cfc86499c8502fb81ee31b03f.jpg",[],{"type":17,"tag":25,"props":234,"children":235},{},[236,238],{"type":23,"value":237},"**2. **超大模型关键特性之——",{"type":17,"tag":67,"props":239,"children":240},{},[241],{"type":23,"value":242},"增量推理（state resuse）及部署",{"type":17,"tag":25,"props":244,"children":245},{},[246],{"type":23,"value":247},"对于自回归（Auto-regressive）的语言模型，随着新词的生成，推理输入长度不断增大。如果使用动态图执行，在不同迭代间，图中每个算子的shape发生改变，无法利用之前缓存的算子编译信息，会影响性能。",{"type":17,"tag":25,"props":249,"children":250},{},[251],{"type":23,"value":252},"我们通过修改推理脚本，实现了增量推理的功能，一次增量推理的流程包含两个阶段：第一阶段推理，输入为全量，将输入padding到固定长度，保存到state，这一阶段是为了处理不定长的输入，将输入的state保存下来；后续推理阶段，输入字长为1，即每次输入上一步生成的token，利用保存的state实现增量推理，产生下一个token并更新state。由于非线性操作的存在，此种增量推理的流程与正常全量推理并不完全等价，不过我们在下游任务推理中发现，增量推理的精度并无明显劣化。我们通过鹏程.盘古在昇腾硬件平台上进行了一系列实验，结果如下图所示：",{"type":17,"tag":25,"props":254,"children":255},{},[256],{"type":17,"tag":29,"props":257,"children":259},{"alt":7,"src":258},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/377c2b4352944380baade73e6499c483.jpg",[],{"type":17,"tag":25,"props":261,"children":262},{},[263],{"type":23,"value":264},"在增量推理过程中，第一阶段只执行一次，而第二阶段会执行多次，比单纯的全量推理，整体性能提升比较明显。",{"type":17,"tag":25,"props":266,"children":267},{},[268],{"type":17,"tag":29,"props":269,"children":271},{"alt":7,"src":270},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/e7921f9f668945aa841314a888615fef.jpg",[],{"type":17,"tag":25,"props":273,"children":274},{},[275],{"type":23,"value":276},"MindSpore Serving支持增量推理模型部署，包括单卡模型和分布式模型场景。使用自定义子图间编排串接两个不同序列长度输入的执行，维护模型的状态，避免多个请求的执行干扰。",{"type":17,"tag":25,"props":278,"children":279},{},[280],{"type":17,"tag":29,"props":281,"children":283},{"alt":7,"src":282},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/b5755940d1114b47b645f6e6628cf4d9.jpg",[],{"type":17,"tag":25,"props":285,"children":286},{},[287],{"type":23,"value":123},{"type":17,"tag":25,"props":289,"children":290},{},[291],{"type":17,"tag":48,"props":292,"children":295},{"href":293,"rel":294},"https://gitee.com/mindspore/docs/blob/r1.3/docs/serving/docs/source%5C_zh%5C_cn/serving%5C_distributed%5C_example.md",[52],[296],{"type":23,"value":297},"https://gitee.com/mindspore/docs/blob/r1.3/docs/serving/docs/source\\_zh\\_cn/serving\\_distributed\\_example.md",{"type":17,"tag":25,"props":299,"children":300},{},[301],{"type":17,"tag":67,"props":302,"children":303},{},[304],{"type":23,"value":305},"MindSpore Lite 端云训练",{"type":17,"tag":25,"props":307,"children":308},{},[309],{"type":23,"value":310},"随着用户对数据隐私要求越来越高，许多用户隐私数据无法上传到云侧，因此在用户端侧进行训练将逐渐成为一个趋势，并且端云协同训练将可以充分利用端侧的计算资源，进一步降低训练所需的时间。当前MindSpore Lite支持对MindSpore训练导出的模型进行增量训练，实现云-端训练的无缝切换。但由于端侧硬件资源的限制，如何在不影响用户体验下进行训练，对端侧训练的内存和性能提出了挑战。",{"type":17,"tag":25,"props":312,"children":313},{},[314,316,321],{"type":23,"value":315},"MindSpore Lite采用了训练内存复用、virtual batch、混合精度训练、在线融合、量化等手段对减少端侧训练时的内存占用。同时在联邦学习MindSpore Federated场景下支持云侧对端侧权重的读写，权重采用差分隐私方式进行端云传输",{"type":17,"tag":67,"props":317,"children":318},{},[319],{"type":23,"value":320},"进一步保证了端云训练中的安全性",{"type":23,"value":322},"。端云训练一般流程如下：",{"type":17,"tag":25,"props":324,"children":325},{},[326],{"type":17,"tag":29,"props":327,"children":329},{"alt":7,"src":328},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/ae9dcdd1be6146e6af72eaea1b556693.jpg",[],{"type":17,"tag":25,"props":331,"children":332},{},[333],{"type":23,"value":334},"端侧训练特性demo试用：",{"type":17,"tag":25,"props":336,"children":337},{},[338],{"type":17,"tag":48,"props":339,"children":342},{"href":340,"rel":341},"https://gitee.com/mindspore/docs/blob/r1.3/docs/lite/docs/source%5C_zh%5C_cn/quick%5C_start/train%5C_lenet.md",[52],[343],{"type":23,"value":344},"https://gitee.com/mindspore/docs/blob/r1.3/docs/lite/docs/source\\_zh\\_cn/quick\\_start/train\\_lenet.md",{"type":17,"tag":25,"props":346,"children":347},{},[348],{"type":17,"tag":67,"props":349,"children":350},{},[351],{"type":23,"value":352},"调试器易用性更上一层楼：图码结合调试和训练回放",{"type":17,"tag":25,"props":354,"children":355},{},[356],{"type":23,"value":357},"作为MindSpore图模式下的调试利器，调试器提供了丰富的检查规则帮助用户快速识别常见精度问题。为了帮助大家更好地在图模式下调试脚本，1.3版本中，我们新增了图码结合调试和训练回放功能。",{"type":17,"tag":359,"props":360,"children":361},"ul",{},[362,368],{"type":17,"tag":363,"props":364,"children":365},"li",{},[366],{"type":23,"value":367},"图码结合调试能帮助您掌握代码和计算图的关系，通过调试器提供的代码信息，您能够更好地理解计算图背后的代码逻辑，提升精度问题分析效率。",{"type":17,"tag":363,"props":369,"children":370},{},[371],{"type":23,"value":372},"训练回放是通过离线调试这一全新的调试模式实现的，在训练结束后，您可以通过离线调试模式对训练过程进行分析，还能对并行训练（单机多卡）中的精度问题进行分析和定位。",{"type":17,"tag":25,"props":374,"children":375},{},[376],{"type":17,"tag":67,"props":377,"children":378},{},[379],{"type":23,"value":380},"1. 图码结合调试：一眼掌握代码和计算图的关系",{"type":17,"tag":25,"props":382,"children":383},{},[384],{"type":23,"value":385},"MindSpore图模式的调试中，如果能方便地找到某行代码所关联的计算图节点，对调试效率将有很大提升。在调试器的帮助下，您可以方便地查看计算图中算子节点和代码的关联关系，无论是以码搜图，还是以图找码，都能在图形界面下快速完成。",{"type":17,"tag":25,"props":387,"children":388},{},[389],{"type":23,"value":390},"以码搜图时，输入想要查找的代码行（例如alexnet.py:52），即可寻找同此行代码关联的计算图节点。如下图所示：",{"type":17,"tag":25,"props":392,"children":393},{},[394],{"type":17,"tag":29,"props":395,"children":397},{"alt":7,"src":396},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/e6843baedb674728b4525114aae3e529.jpg",[],{"type":17,"tag":25,"props":399,"children":400},{},[401],{"type":23,"value":402},"以图找码时，在计算图中选中算子节点（例如Conv2D-op108），即可在堆栈信息中查看该算子对应的代码行。如下图所示：",{"type":17,"tag":25,"props":404,"children":405},{},[406],{"type":17,"tag":29,"props":407,"children":409},{"alt":7,"src":408},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/39a0f5a9a2664cf0aa9dbfa77ef5d9d9.jpg",[],{"type":17,"tag":25,"props":411,"children":412},{},[413],{"type":17,"tag":67,"props":414,"children":415},{},[416],{"type":23,"value":417},"2. 支持离线调试：随时离线回放分析训练过程，节约AI处理器资源",{"type":17,"tag":25,"props":419,"children":420},{},[421],{"type":23,"value":422},"在使用之前的调试器时，要求先启动训练进程，在训练进程运行的过程中对精度问题进行分析。我们提供了离线调试的能力，通过dump功能将调试数据保存到磁盘上，然后就可以在不使用昇腾AI处理器的情况下回看训练过程，分析精度问题。由于数据在磁盘上，离线调试中还可以随意切换正在调试的迭代，回放训练，免去在线调试错过关键迭代后又要从头运行脚本的烦恼",{"type":17,"tag":25,"props":424,"children":425},{},[426],{"type":17,"tag":67,"props":427,"children":428},{},[429],{"type":23,"value":430},"3. 支持单机多卡调试：定位并行训练中的精度问题",{"type":17,"tag":25,"props":432,"children":433},{},[434],{"type":23,"value":435},"在离线调试功能的基础上，调试器支持了单机多卡训练的调试。在MindSpore上调试单机多卡的训练时，只需通过dump功能保存数据到磁盘，就可以使用MindInsight可视化地进行分析。调试器中已经提供的监测点，例如检查梯度消失，检查激活值饱和，检查权重不更新等，都可以继续在此场景下使用。图码结合调试同样支持单机多卡。",{"type":17,"tag":25,"props":437,"children":438},{},[439],{"type":23,"value":440},"查看教程：",{"type":17,"tag":25,"props":442,"children":443},{},[444],{"type":17,"tag":48,"props":445,"children":448},{"href":446,"rel":447},"https://gitee.com/mindspore/docs/blob/r1.3/docs/mindinsight/docs/source%5C_zh%5C_cn/debugger%5C_online.md",[52],[449],{"type":23,"value":450},"https://gitee.com/mindspore/docs/blob/r1.3/docs/mindinsight/docs/source\\_zh\\_cn/debugger\\_online.md",{"type":17,"tag":25,"props":452,"children":453},{},[454],{"type":17,"tag":48,"props":455,"children":458},{"href":456,"rel":457},"https://gitee.com/mindspore/docs/blob/r1.3/docs/mindinsight/docs/source%5C_zh%5C_cn/debugger%5C_offline.md",[52],[459],{"type":23,"value":460},"https://gitee.com/mindspore/docs/blob/r1.3/docs/mindinsight/docs/source\\_zh\\_cn/debugger\\_offline.md",{"type":17,"tag":25,"props":462,"children":463},{},[464],{"type":17,"tag":67,"props":465,"children":466},{},[467],{"type":23,"value":468},"推理优化——X86_64 CPU PC推理能力",{"type":17,"tag":25,"props":470,"children":471},{},[472,474,479,481,486],{"type":23,"value":473},"为了更好的支持PC侧推理，x86_64从汇编层面入手，支持针对不同卷积shape的动态block切分，充分利用寄存器等硬件资源，使我们推理性能达到极致，较上个版本推理时延",{"type":17,"tag":67,"props":475,"children":476},{},[477],{"type":23,"value":478},"有了10%~75%+的提升",{"type":23,"value":480},"。我们在Intel Core i7-8700 CPU上与OpenVINO（2021.3.394）、MNN（1.2.0）、TNN（v0.3）在几个经典CV类网络上进行benchmark测试，从测试结果可以看出MindSpore Lite",{"type":17,"tag":67,"props":482,"children":483},{},[484],{"type":23,"value":485},"保持了较高的水平",{"type":23,"value":94},{"type":17,"tag":25,"props":488,"children":489},{},[490],{"type":17,"tag":29,"props":491,"children":493},{"alt":7,"src":492},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/6de700a235b548e4a7e0a1c444f47903.jpg",[],{"type":17,"tag":25,"props":495,"children":496},{},[497],{"type":17,"tag":67,"props":498,"children":499},{},[500],{"type":23,"value":501},"更多推理性能优化",{"type":17,"tag":25,"props":503,"children":504},{},[505],{"type":23,"value":506},"随着数据集和网络规模越来越大，网络计算量也越来越大；同时交互式推理任务中的对时延要求更加严格，深度神经网络推理任务逐渐向AI加速硬件(比如GPU)进行迁移。MindSpore 1.3版本提供了更多与此相关的推理性能进行优化，性能相比此前大幅提升。",{"type":17,"tag":25,"props":508,"children":509},{},[510,512,517],{"type":23,"value":511},"基于Model Zoo中的一些典型的网络，分别使用MindSpore1.2和MindSpore1.3版本对，统计推理请求平均执行时间，",{"type":17,"tag":67,"props":513,"children":514},{},[515],{"type":23,"value":516},"推理性能提升3~8倍",{"type":23,"value":94},{"type":17,"tag":25,"props":519,"children":520},{},[521],{"type":17,"tag":29,"props":522,"children":524},{"alt":7,"src":523},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/a5c4aaf5b5aa46ca9c89e3bb00b5046c.jpg",[],{"type":17,"tag":25,"props":526,"children":527},{},[528],{"type":17,"tag":29,"props":529,"children":531},{"alt":7,"src":530},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/7239edee5f674658a6c51e037ed5759c.jpg",[],{"type":17,"tag":25,"props":533,"children":534},{},[535],{"type":17,"tag":67,"props":536,"children":537},{},[538],{"type":23,"value":539},"图算融合加速MindSpore网络训练",{"type":17,"tag":25,"props":541,"children":542},{},[543,545,550],{"type":23,"value":544},"图算融合是MindSpore的关键技术之一，通过图层融合与算子生成的协同优化来提升网络执行性能。在之前版本，我们使能了NPU（昇腾）和GPU上图算融合的基本能力，并且在标杆网络上取得了不错的成绩。在1.3版本中，我们加强了图算融合在GPU上的泛化能力，通过对Model Zoo 40多张主流网络的验证，平均可",{"type":17,"tag":67,"props":546,"children":547},{},[548],{"type":23,"value":549},"获得89%的性能提升",{"type":23,"value":551},"。以Transformer为例，使能图算后，从2.5小时/epoch降低到1.75小时/epoch，完整训练时间从5.4天降低到3.8天！",{"type":17,"tag":25,"props":553,"children":554},{},[555],{"type":23,"value":556},"同时，在易用性方面我们新增了环境变量的控制方式，使得用户不需要对网络代码做任何侵入修改，即可享受图算融合带来的性能收益：",{"type":17,"tag":25,"props":558,"children":559},{},[560],{"type":17,"tag":561,"props":562,"children":563},"em",{},[564],{"type":23,"value":565},"export MS_GRAPH_KERNEL_FLAGS=”–opt_level=2”",{"type":17,"tag":25,"props":567,"children":568},{},[569],{"type":23,"value":570},"我们选取了部分主流的网络（NLP、推荐及CV）开展性能对比评测，使用图算融合的有普遍的性能提升，提升详情如下图所示：",{"type":17,"tag":25,"props":572,"children":573},{},[574],{"type":17,"tag":29,"props":575,"children":577},{"alt":7,"src":576},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/c48c2ad0989945269f8afdff888e34cb.jpg",[],{"type":17,"tag":25,"props":579,"children":580},{},[581],{"type":23,"value":582},"MindSpore是华为开源的一款全场景AI计算框架，旨在提供友好设计、高效运行、简捷部署的开发体验，目前应用于医疗、金融、科研等多个领域，提供面向端边云多种场景的主流硬件支持，并针对昇腾硬件平台提供深度优化能力。MindSpore着力构筑面向全球的人工智能开源社区，推动人工智能软硬件应用生态繁荣发展。对MindSpore感兴趣的小伙伴扫描下方二维码关注我们吧~",{"type":17,"tag":25,"props":584,"children":585},{},[586],{"type":17,"tag":29,"props":587,"children":589},{"alt":7,"src":588},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/07/16/acea6d4bbc0e44d8bc0f29ced86b5771.jpg",[],{"title":7,"searchDepth":591,"depth":591,"links":592},4,[],"markdown","content:version-updates:zh:653.md","content","version-updates/zh/653.md","version-updates/zh/653","md",1776506145777]