[{"data":1,"prerenderedAt":368},["ShallowReactive",2],{"content-query-vrLHGpYa4Z":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":362,"_id":363,"_source":364,"_file":365,"_stem":366,"_extension":367},"/news/zh/2026-1-14","zh",false,"","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","采用自主创新的「自回归＋扩散解码器」混合架构","2026-1-14","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/07/25/199b735845bf4106b44b2035dc97bd39.png","news",{"type":14,"children":15,"toc":359},"root",[16,24,30,39,47,52,63,83,91,99,104,112,117,125,130,138,143,151,159,164,172,177,184,192,197,204,212,217,224,232,237,244,252,257,264,272,280,285,296,301,310,318,326,331,336,345,350],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"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",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":29},"今天，智谱图像生成模型GLM-Image 正式发布。该模型采用自主创新的「自回归＋扩散解码器」混合架构，是智谱面向「认知型生成」技术范式的一次重要探索，也是首个开源的工业表现级离散自回归图像生成模型。昇腾NPU和昇思MindSpore AI框架提供从数据到训练的全流程支撑，并提出一系列训练性能优化特性，助力智谱高性能训练出首个基于自主创新算力底座全程训练的SOTA模型。",{"type":17,"tag":25,"props":31,"children":32},{},[33],{"type":17,"tag":34,"props":35,"children":36},"strong",{},[37],{"type":23,"value":38},"# 01",{"type":17,"tag":25,"props":40,"children":41},{},[42],{"type":17,"tag":34,"props":43,"children":44},{},[45],{"type":23,"value":46},"GLM-Image模型采用创新的架构实现开源SOTA",{"type":17,"tag":25,"props":48,"children":49},{},[50],{"type":23,"value":51},"GLM-Image模型基于创新的混合架构，实现了在文字渲染的权威榜单中达到开源SOTA水平。",{"type":17,"tag":53,"props":54,"children":56},"div",{"style":55},"text-align: center;",[57],{"type":17,"tag":58,"props":59,"children":62},"img",{"src":60,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/1.jpg","display: block;margin: 0 auto;max-width:70%",[],{"type":17,"tag":64,"props":65,"children":66},"ul",{},[67,73,78],{"type":17,"tag":68,"props":69,"children":70},"li",{},[71],{"type":23,"value":72},"架构革新，面向「认知型生成」的技术探索：GLM-Image采用创新的「自回归 + 扩散编码器」混合架构，兼顾全局指令理解与局部细节刻画，克服了海报、PPT、科普图等知识密集型场景生成难题，向探索类”Nano Banana Pro”的新一代“知识+推理”生成模型迈出了重要一步。",{"type":17,"tag":68,"props":74,"children":75},{},[76],{"type":23,"value":77},"文字渲染开源SOTA：在CVTG-2K（复杂视觉文本生成）和LongText-Bench（长文本渲染）榜单获得开源第一，尤其擅长汉字生成任务。",{"type":17,"tag":68,"props":79,"children":80},{},[81],{"type":23,"value":82},"首个基于自主创新算力底座全程训练的SOTA模型：模型的自回归结构基于昇腾Atlas 800T A2设备与昇思MindSpore AI框架，完成了从数据预处理到大规模训练的全流程构建，验证了在全栈自主创新算力底座上也能训练出SOTA模型。",{"type":17,"tag":25,"props":84,"children":85},{},[86],{"type":17,"tag":34,"props":87,"children":88},{},[89],{"type":23,"value":90},"# 02",{"type":17,"tag":25,"props":92,"children":93},{},[94],{"type":17,"tag":34,"props":95,"children":96},{},[97],{"type":23,"value":98},"昇腾+昇思MindSpore支持GLM-Image模型端到端高效训练",{"type":17,"tag":25,"props":100,"children":101},{},[102],{"type":23,"value":103},"GLM-Image系列模型需要处理大量文字和图片数据，对模型训练的内存开销和性能提出了更高的挑战。为了应对这些挑战，昇腾+昇思MindSpore引入动态图多级流水下发、多流并行执行、高性能融合算子等特性，全面提升从数据处理、预训练、SFT到RL的端到端性能。",{"type":17,"tag":25,"props":105,"children":106},{},[107],{"type":17,"tag":34,"props":108,"children":109},{},[110],{"type":23,"value":111},"1、支持动态图多级流水下发，大幅加速Host侧算子下发",{"type":17,"tag":25,"props":113,"children":114},{},[115],{"type":23,"value":116},"NPU+CPU服务器常配备大规模多核CPU资源，单线程的算子下发往往跟不上NPU的计算，Host侧算力难以充分发挥，导致Device侧空转与整体资源利用率偏低。因此，使用了多级流水优化机制，依托昇腾+昇思MindSpore的软硬件协同，将Python执行、算子shape推导与算子下发等关键阶段进行流水化并高度重叠，有效提升Host侧并行度与下发效率，提升训练性能20%。",{"type":17,"tag":25,"props":118,"children":119},{},[120],{"type":17,"tag":34,"props":121,"children":122},{},[123],{"type":23,"value":124},"2、支持多流并行执行，打破通信墙",{"type":17,"tag":25,"props":126,"children":127},{},[128],{"type":23,"value":129},"多模态训练场景中，文本梯度同步、图像特征广播、混合并行等操作都有复杂的通信需求。单流串行执行时，计算和通信两类操作互相等待，整体资源利用率低。因此，通过多流并行执行机制，为通信和计算分别配置通信流和计算流，两者公用内存池来实现内存复用，允许通信和计算两类操作同时运行。此外，不同通信域还能单独申请和分配通信流，充分利用网络带宽消除通信串扰，整体提升训练性能10%。",{"type":17,"tag":25,"props":131,"children":132},{},[133],{"type":17,"tag":34,"props":134,"children":135},{},[136],{"type":23,"value":137},"3、昇腾亲和的高性能融合算子，提升训练性能和稳定性",{"type":17,"tag":25,"props":139,"children":140},{},[141],{"type":23,"value":142},"昇腾CANN高性能算子库提供多种高性能融合算子：AdamW EMA融合算子，缓解扩散模型训练后期loss震荡问题，使收敛效果更稳定；COC通算融合算子，使用集合通信创新算法，用计算掩盖多机多卡场景TP域中ReduceScatter和AllGather等核心通信算子的通信开销，通信效率提升15%。",{"type":17,"tag":25,"props":144,"children":145},{},[146],{"type":17,"tag":34,"props":147,"children":148},{},[149],{"type":23,"value":150},"# 03",{"type":17,"tag":25,"props":152,"children":153},{},[154],{"type":17,"tag":34,"props":155,"children":156},{},[157],{"type":23,"value":158},"基于昇腾+昇思MindSpore的GLM-Image模型的多模态体验",{"type":17,"tag":25,"props":160,"children":161},{},[162],{"type":23,"value":163},"让我们来看看 GLM-Image 在实际的复杂图文等任务中的表现。",{"type":17,"tag":25,"props":165,"children":166},{},[167],{"type":17,"tag":34,"props":168,"children":169},{},[170],{"type":23,"value":171},"场景一：科普插画",{"type":17,"tag":25,"props":173,"children":174},{},[175],{"type":23,"value":176},"GLM-Image 更擅长绘制包含复杂逻辑流程与文字说明的科普插画及原理示意图。",{"type":17,"tag":53,"props":178,"children":179},{"style":55},[180],{"type":17,"tag":58,"props":181,"children":183},{"src":182,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/2.jpg",[],{"type":17,"tag":25,"props":185,"children":186},{},[187],{"type":17,"tag":34,"props":188,"children":189},{},[190],{"type":23,"value":191},"场景二：多格图画",{"type":17,"tag":25,"props":193,"children":194},{},[195],{"type":23,"value":196},"在生成电商图、漫画等多格图画时，GLM-Image能够保持风格和主体的一致性，并保障多处文字生成的准确率。",{"type":17,"tag":53,"props":198,"children":199},{"style":55},[200],{"type":17,"tag":58,"props":201,"children":203},{"src":202,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/3.jpg",[],{"type":17,"tag":25,"props":205,"children":206},{},[207],{"type":17,"tag":34,"props":208,"children":209},{},[210],{"type":23,"value":211},"场景三：社交媒体图文",{"type":17,"tag":25,"props":213,"children":214},{},[215],{"type":23,"value":216},"GLM-Image 适用于制作社交媒体封面及内容等排版复杂的图片，让您的创作更自由丰富。",{"type":17,"tag":53,"props":218,"children":219},{"style":55},[220],{"type":17,"tag":58,"props":221,"children":223},{"src":222,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/4.jpg",[],{"type":17,"tag":25,"props":225,"children":226},{},[227],{"type":17,"tag":34,"props":228,"children":229},{},[230],{"type":23,"value":231},"场景四：商业海报",{"type":17,"tag":25,"props":233,"children":234},{},[235],{"type":23,"value":236},"GLM-Image 能够生成构图富有设计感、文字嵌入准确的节日海报与商业宣传图。",{"type":17,"tag":53,"props":238,"children":239},{"style":55},[240],{"type":17,"tag":58,"props":241,"children":243},{"src":242,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/5.jpg",[],{"type":17,"tag":25,"props":245,"children":246},{},[247],{"type":17,"tag":34,"props":248,"children":249},{},[250],{"type":23,"value":251},"场景五：写实摄影",{"type":17,"tag":25,"props":253,"children":254},{},[255],{"type":23,"value":256},"在文字渲染以外，GLM-Image也同样擅长生成各种景别和尺寸的人像、宠物、风景、静物。",{"type":17,"tag":53,"props":258,"children":259},{"style":55},[260],{"type":17,"tag":58,"props":261,"children":263},{"src":262,"style":61,"alt":7},"/category/information/news/banner/2026-1-14/6.jpg",[],{"type":17,"tag":25,"props":265,"children":266},{},[267],{"type":17,"tag":34,"props":268,"children":269},{},[270],{"type":23,"value":271},"# 04",{"type":17,"tag":25,"props":273,"children":274},{},[275],{"type":17,"tag":34,"props":276,"children":277},{},[278],{"type":23,"value":279},"快速上手 基于昇腾玩转GLM-Image模型推理",{"type":17,"tag":25,"props":281,"children":282},{},[283],{"type":23,"value":284},"体验指导：",{"type":17,"tag":25,"props":286,"children":287},{},[288],{"type":17,"tag":289,"props":290,"children":294},"a",{"href":291,"rel":292},"https://modelers.cn/models/MindIE/GLM-Image",[293],"nofollow",[295],{"type":23,"value":291},{"type":17,"tag":25,"props":297,"children":298},{},[299],{"type":23,"value":300},"镜像链接：",{"type":17,"tag":25,"props":302,"children":303},{},[304],{"type":17,"tag":289,"props":305,"children":308},{"href":306,"rel":307},"https://mindie.obs.cn-north-4.myhuaweicloud.com/zhipu/mindie_2.2.RC1-800I-A2-Glm-Image-py311-openeuler24.03-lts-aarch64.tar.gz",[293],[309],{"type":23,"value":306},{"type":17,"tag":25,"props":311,"children":312},{},[313],{"type":17,"tag":34,"props":314,"children":315},{},[316],{"type":23,"value":317},"# 05",{"type":17,"tag":25,"props":319,"children":320},{},[321],{"type":17,"tag":34,"props":322,"children":323},{},[324],{"type":23,"value":325},"结语",{"type":17,"tag":25,"props":327,"children":328},{},[329],{"type":23,"value":330},"智谱GLM系列模型会持续迭代，昇思MindSpore也将同步上线更加丰富的特性，欢迎关注昇思MindSpore社区。",{"type":17,"tag":25,"props":332,"children":333},{},[334],{"type":23,"value":335},"昇思MindSpore开源社区：",{"type":17,"tag":25,"props":337,"children":338},{},[339],{"type":17,"tag":289,"props":340,"children":343},{"href":341,"rel":342},"https://www.mindspore.cn/",[293],[344],{"type":23,"value":341},{"type":17,"tag":25,"props":346,"children":347},{},[348],{"type":23,"value":349},"昇思MindSpore代码仓库：",{"type":17,"tag":25,"props":351,"children":352},{},[353],{"type":17,"tag":289,"props":354,"children":357},{"href":355,"rel":356},"https://atomgit.com/mindspore",[293],[358],{"type":23,"value":355},{"title":7,"searchDepth":360,"depth":360,"links":361},4,[],"markdown","content:news:zh:2026-1-14.md","content","news/zh/2026-1-14.md","news/zh/2026-1-14","md",1776506060880]