[{"data":1,"prerenderedAt":683},["ShallowReactive",2],{"content-query-CoDD6I70TP":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":677,"_id":678,"_source":679,"_file":680,"_stem":681,"_extension":682},"/technology-blogs/zh/3709","zh",false,"","0Day支持！昇思MindSpore同步首发Qwen3，支持一键部署","Qwen3是阿里云于2025年4月29日发布并开源的全新模型，作为Qwen 系列中的最新一代大型语言模型，提供了一系列密集型和混合专家（MoE）模型。","2025-04-29","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/04/30/612e020359004152a095b8072254b984.png","technology-blogs","实践",{"type":15,"children":16,"toc":667},"root",[17,25,31,36,41,46,51,56,61,72,77,86,91,100,105,114,119,124,133,138,147,152,161,166,175,180,189,194,199,208,213,222,240,251,265,280,288,296,301,309,317,343,357,362,377,382,400,408,423,436,441,461,474,484,492,497,505,510,518,523,531,536,544,549,557,567,584,589,597,614,619,627,644,649,657,662],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"_0day支持昇思mindspore同步首发qwen3支持一键部署",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":24,"value":30},"Qwen3是阿里云于2025年4月29日发布并开源的全新模型，作为Qwen 系列中的最新一代大型语言模型，提供了一系列密集型和混合专家（MoE）模型。本次Qwen发布多个尺寸模型，覆盖235B/32B/30B/14B/8B/4B/1.7B/0.6B。昇思MindSpore基于对Qwen2.5的支持与兼容主流生态的接口，快速实现Qwen3的0Day支持，并将MindSpore版Qwen3代码上传至开源社区代码仓，面向开发者提供开箱即用的模型。",{"type":18,"tag":26,"props":32,"children":33},{},[34],{"type":24,"value":35},"模型链接：",{"type":18,"tag":26,"props":37,"children":38},{},[39],{"type":24,"value":40},"类型",{"type":18,"tag":26,"props":42,"children":43},{},[44],{"type":24,"value":45},"模型名称",{"type":18,"tag":26,"props":47,"children":48},{},[49],{"type":24,"value":50},"魔乐社区链接",{"type":18,"tag":26,"props":52,"children":53},{},[54],{"type":24,"value":55},"稠密Base",{"type":18,"tag":26,"props":57,"children":58},{},[59],{"type":24,"value":60},"Qwen3-0.6B-Base",{"type":18,"tag":26,"props":62,"children":63},{},[64],{"type":18,"tag":65,"props":66,"children":70},"a",{"href":67,"rel":68},"https://modelers.cn/models/MindSpore-Lab/Qwen3-0.6B-Base",[69],"nofollow",[71],{"type":24,"value":67},{"type":18,"tag":26,"props":73,"children":74},{},[75],{"type":24,"value":76},"Qwen3-1.7B-Base",{"type":18,"tag":26,"props":78,"children":79},{},[80],{"type":18,"tag":65,"props":81,"children":84},{"href":82,"rel":83},"https://modelers.cn/models/MindSpore-Lab/Qwen3-1.7B-Base",[69],[85],{"type":24,"value":82},{"type":18,"tag":26,"props":87,"children":88},{},[89],{"type":24,"value":90},"Qwen3-4B-Base",{"type":18,"tag":26,"props":92,"children":93},{},[94],{"type":18,"tag":65,"props":95,"children":98},{"href":96,"rel":97},"https://modelers.cn/models/MindSpore-Lab/Qwen3-4B-Base",[69],[99],{"type":24,"value":96},{"type":18,"tag":26,"props":101,"children":102},{},[103],{"type":24,"value":104},"Qwen3-8B-Base",{"type":18,"tag":26,"props":106,"children":107},{},[108],{"type":18,"tag":65,"props":109,"children":112},{"href":110,"rel":111},"https://modelers.cn/models/MindSpore-Lab/Qwen3-8B-Base",[69],[113],{"type":24,"value":110},{"type":18,"tag":26,"props":115,"children":116},{},[117],{"type":24,"value":118},"稠密Instruct",{"type":18,"tag":26,"props":120,"children":121},{},[122],{"type":24,"value":123},"Qwen3-0.6B",{"type":18,"tag":26,"props":125,"children":126},{},[127],{"type":18,"tag":65,"props":128,"children":131},{"href":129,"rel":130},"https://modelers.cn/models/MindSpore-Lab/Qwen3-0.6B",[69],[132],{"type":24,"value":129},{"type":18,"tag":26,"props":134,"children":135},{},[136],{"type":24,"value":137},"Qwen3-1.7B",{"type":18,"tag":26,"props":139,"children":140},{},[141],{"type":18,"tag":65,"props":142,"children":145},{"href":143,"rel":144},"https://modelers.cn/models/MindSpore-Lab/Qwen3-1.7B",[69],[146],{"type":24,"value":143},{"type":18,"tag":26,"props":148,"children":149},{},[150],{"type":24,"value":151},"Qwen3-4B",{"type":18,"tag":26,"props":153,"children":154},{},[155],{"type":18,"tag":65,"props":156,"children":159},{"href":157,"rel":158},"https://modelers.cn/models/MindSpore-Lab/Qwen3-4B",[69],[160],{"type":24,"value":157},{"type":18,"tag":26,"props":162,"children":163},{},[164],{"type":24,"value":165},"Qwen3-8B",{"type":18,"tag":26,"props":167,"children":168},{},[169],{"type":18,"tag":65,"props":170,"children":173},{"href":171,"rel":172},"https://modelers.cn/models/MindSpore-Lab/Qwen3-8B",[69],[174],{"type":24,"value":171},{"type":18,"tag":26,"props":176,"children":177},{},[178],{"type":24,"value":179},"Qwen3-32B",{"type":18,"tag":26,"props":181,"children":182},{},[183],{"type":18,"tag":65,"props":184,"children":187},{"href":185,"rel":186},"https://modelers.cn/models/MindSpore-Lab/Qwen3-32B",[69],[188],{"type":24,"value":185},{"type":18,"tag":26,"props":190,"children":191},{},[192],{"type":24,"value":193},"稀疏MOE",{"type":18,"tag":26,"props":195,"children":196},{},[197],{"type":24,"value":198},"Qwen3-30B-A3B",{"type":18,"tag":26,"props":200,"children":201},{},[202],{"type":18,"tag":65,"props":203,"children":206},{"href":204,"rel":205},"https://modelers.cn/models/MindSpore-Lab/Qwen3-30B-A3B",[69],[207],{"type":24,"value":204},{"type":18,"tag":26,"props":209,"children":210},{},[211],{"type":24,"value":212},"Qwen3-235B-A22B",{"type":18,"tag":26,"props":214,"children":215},{},[216],{"type":18,"tag":65,"props":217,"children":220},{"href":218,"rel":219},"https://modelers.cn/models/MindSpore-Lab/Qwen3-235B-A22B",[69],[221],{"type":24,"value":218},{"type":18,"tag":223,"props":224,"children":226},"h3",{"id":225},"_01-qwen3模型介绍",[227,233,235],{"type":18,"tag":228,"props":229,"children":230},"strong",{},[231],{"type":24,"value":232},"# 01",{"type":24,"value":234}," ",{"type":18,"tag":228,"props":236,"children":237},{},[238],{"type":24,"value":239},"Qwen3模型介绍",{"type":18,"tag":26,"props":241,"children":242},{},[243,245,249],{"type":24,"value":244},"Qwen3是 Qwen 系列大型语言模型的最新成员。其中的旗舰模型 ",{"type":18,"tag":228,"props":246,"children":247},{},[248],{"type":24,"value":212},{"type":24,"value":250}," 在代码、数学、通用能力等基准测试中，与 DeepSeek-R1、o1、o3-mini、Grok-3 和 Gemini-2.5-Pro 等顶级模型相比，表现出极具竞争力的结果。此外，小型 MoE 模型 Qwen3-30B-A3B 的激活参数数量是 QwQ-32B 的 10%，表现更胜一筹，甚至像 Qwen3-4B 这样的小模型也能匹敌 Qwen2.5-72B-Instruct 的性能。",{"type":18,"tag":26,"props":252,"children":253},{},[254,259,260],{"type":18,"tag":228,"props":255,"children":256},{},[257],{"type":24,"value":258},"1",{"type":24,"value":234},{"type":18,"tag":228,"props":261,"children":262},{},[263],{"type":24,"value":264},"Qwen3 模型支持两种思考模式",{"type":18,"tag":266,"props":267,"children":268},"ul",{},[269,275],{"type":18,"tag":270,"props":271,"children":272},"li",{},[273],{"type":24,"value":274},"**思考模式：**在这种模式下，模型会逐步推理，经过深思熟虑后给出最终答案。这种方法适合需要深入思考的复杂问题。",{"type":18,"tag":270,"props":276,"children":277},{},[278],{"type":24,"value":279},"**非思考模式：**在此模式中，模型提供快速、近乎即时的响应，适用于那些对速度要求高于深度的简单问题。",{"type":18,"tag":26,"props":281,"children":282},{},[283],{"type":18,"tag":228,"props":284,"children":285},{},[286],{"type":24,"value":287},"2",{"type":18,"tag":26,"props":289,"children":290},{},[291],{"type":18,"tag":228,"props":292,"children":293},{},[294],{"type":24,"value":295},"多语言",{"type":18,"tag":26,"props":297,"children":298},{},[299],{"type":24,"value":300},"Qwen3 模型支持 119 种语言和方言。这一广泛的多语言能力为国际应用开辟了新的可能性。",{"type":18,"tag":26,"props":302,"children":303},{},[304],{"type":18,"tag":228,"props":305,"children":306},{},[307],{"type":24,"value":308},"3",{"type":18,"tag":26,"props":310,"children":311},{},[312],{"type":18,"tag":228,"props":313,"children":314},{},[315],{"type":24,"value":316},"预训练",{"type":18,"tag":26,"props":318,"children":319},{},[320,322,327,329,334,336,341],{"type":24,"value":321},"在预训练方面，Qwen3 的数据集相比 Qwen2.5 有了显著扩展。Qwen2.5是在 ",{"type":18,"tag":228,"props":323,"children":324},{},[325],{"type":24,"value":326},"18 万亿个 token",{"type":24,"value":328}," 上进行预训练的，而 Qwen3 使用的数据量几乎是其两倍，达到了约 ",{"type":18,"tag":228,"props":330,"children":331},{},[332],{"type":24,"value":333},"36 万亿个 token",{"type":24,"value":335},"，涵盖了 ",{"type":18,"tag":228,"props":337,"children":338},{},[339],{"type":24,"value":340},"119 种语言和方言",{"type":24,"value":342},"。",{"type":18,"tag":26,"props":344,"children":345},{},[346,351,352],{"type":18,"tag":228,"props":347,"children":348},{},[349],{"type":24,"value":350},"4",{"type":24,"value":234},{"type":18,"tag":228,"props":353,"children":354},{},[355],{"type":24,"value":356},"后训练",{"type":18,"tag":26,"props":358,"children":359},{},[360],{"type":24,"value":361},"Qwen3实施了一个四阶段的训练流程。该流程包括：（1）长思维链冷启动，（2）长思维链强化学习，（3）思维模式融合，以及（4）通用强化学习。",{"type":18,"tag":223,"props":363,"children":365},{"id":364},"_02-昇思mindspore相关技术特性",[366,371,372],{"type":18,"tag":228,"props":367,"children":368},{},[369],{"type":24,"value":370},"# 02",{"type":24,"value":234},{"type":18,"tag":228,"props":373,"children":374},{},[375],{"type":24,"value":376},"昇思MindSpore相关技术特性",{"type":18,"tag":26,"props":378,"children":379},{},[380],{"type":24,"value":381},"昇思MindSpore原生支持Qwen系列大模型，兼容主流生态的分布式并行接口，已快速完成Qwen3多个模型的同步支持。同时，具备以下技术特性，加速模型训练、推理。",{"type":18,"tag":266,"props":383,"children":384},{},[385,390,395],{"type":18,"tag":270,"props":386,"children":387},{},[388],{"type":24,"value":389},"昇思MindSpore同步支持MindSpeed加速库，已完成Qwen3系列模型训练适配。",{"type":18,"tag":270,"props":391,"children":392},{},[393],{"type":24,"value":394},"**JIT加速，提升推理系统吞吐率：**昇思MindSpore通过JIT编译自动将模型的Python类或者函数，编译成一张完整的计算图，进而通过自动模式匹配，在整图范围内将多种小算子组合，融合成单个大颗粒的算子。同时，构建了Shape推导、Tiling数据计算、下发执行的三级流水线，实现Host计算和Device计算的掩盖，有效提升了计算图动态Shape执行效率。",{"type":18,"tag":270,"props":396,"children":397},{},[398],{"type":24,"value":399},"**无缝接入vLLM生态：**昇思MindSpore开发了vLLM-MindSpore插件，支持主流大模型的推理服务部署。如图1所示，其采用MSAdapter将vLLM服务组件依赖的PyTorch接口映射至MindSpore能力，无缝继承了Continuous Batching等核心特性。通过插件式接入MindSpore Transformers套件所提供的大模型，实现昇腾+昇思的推理加速，有机整合了vLLM和MindSpore的推理加速能力。vLLM-MindSpore插件已适配vLLM v0.7.3版本，即将支持v0.8.3版本和V1架构。",{"type":18,"tag":26,"props":401,"children":402},{},[403],{"type":18,"tag":404,"props":405,"children":407},"img",{"alt":7,"src":406},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/04/30/932b09217c4c46ee86424abf6515fc05.png",[],{"type":18,"tag":223,"props":409,"children":411},{"id":410},"_03-手把手教程基于昇思快速上手qwen3模型训练和推理",[412,417,418],{"type":18,"tag":228,"props":413,"children":414},{},[415],{"type":24,"value":416},"# 03",{"type":24,"value":234},{"type":18,"tag":228,"props":419,"children":420},{},[421],{"type":24,"value":422},"手把手教程：基于昇思快速上手Qwen3模型训练和推理",{"type":18,"tag":26,"props":424,"children":425},{},[426,430,431],{"type":18,"tag":228,"props":427,"children":428},{},[429],{"type":24,"value":258},{"type":24,"value":234},{"type":18,"tag":228,"props":432,"children":433},{},[434],{"type":24,"value":435},"快速开始",{"type":18,"tag":26,"props":437,"children":438},{},[439],{"type":24,"value":440},"以Qwen3-32B推理为例，使用1台（2卡）Atlas 800I A2（64G）服务器（基于BF16权重）。昇思MindSpore提供了Qwen3-32B推理专用的Docker容器镜像，供开发者快速体验。",{"type":18,"tag":26,"props":442,"children":443},{},[444,446,451,453],{"type":24,"value":445},"下载昇思 ",{"type":18,"tag":228,"props":447,"children":448},{},[449],{"type":24,"value":450},"MindSpore Qwen3",{"type":24,"value":452},"****-32B**** ",{"type":18,"tag":228,"props":454,"children":455},{},[456],{"type":18,"tag":228,"props":457,"children":458},{},[459],{"type":24,"value":460},"推理容器镜像",{"type":18,"tag":26,"props":462,"children":463},{},[464],{"type":18,"tag":228,"props":465,"children":466},{},[467,472],{"type":18,"tag":228,"props":468,"children":469},{},[470],{"type":24,"value":471},"执行以下",{"type":24,"value":473}," Shell 命令，拉取昇思 MindSpore Qwen3 推理容器镜像：",{"type":18,"tag":475,"props":476,"children":478},"pre",{"code":477},"\ndocker pull swr.cn-central-221.ovaijisuan.com/mindformers/qwen3_mindspore2.6.0-infer:20250428\n",[479],{"type":18,"tag":480,"props":481,"children":482},"code",{"__ignoreMap":7},[483],{"type":24,"value":477},{"type":18,"tag":26,"props":485,"children":486},{},[487],{"type":18,"tag":228,"props":488,"children":489},{},[490],{"type":24,"value":491},"启动容器",{"type":18,"tag":26,"props":493,"children":494},{},[495],{"type":24,"value":496},"执行以下命令创建并启动容器：",{"type":18,"tag":475,"props":498,"children":500},{"code":499},"docker run -it --privileged --name=qwen3 --net=host \\\n   --shm-size 500g \\\n   --device=/dev/davinci0 \\\n   --device=/dev/davinci1 \\\n   --device=/dev/davinci2 \\\n   --device=/dev/davinci3 \\\n   --device=/dev/davinci4 \\\n   --device=/dev/davinci5 \\\n   --device=/dev/davinci6 \\\n   --device=/dev/davinci7 \\\n   --device=/dev/davinci_manager \\\n   --device=/dev/hisi_hdc \\\n   --device /dev/devmm_svm \\\n   -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \\\n   -v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \\\n   -v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \\\n   -v /usr/local/sbin:/usr/local/sbin \\\n   -v /etc/hccn.conf:/etc/hccn.conf \\\n   swr.cn-central-221.ovaijisuan.com/mindformers/qwen3_mindspore2.6.0-infer:20250428 \\\n   bash\n",[501],{"type":18,"tag":480,"props":502,"children":503},{"__ignoreMap":7},[504],{"type":24,"value":499},{"type":18,"tag":26,"props":506,"children":507},{},[508],{"type":24,"value":509},"后续所有操作均在容器内操作。",{"type":18,"tag":26,"props":511,"children":512},{},[513],{"type":18,"tag":228,"props":514,"children":515},{},[516],{"type":24,"value":517},"模型下载",{"type":18,"tag":26,"props":519,"children":520},{},[521],{"type":24,"value":522},"执行以下命令为自定义下载路径/home/work添加白名单：",{"type":18,"tag":475,"props":524,"children":526},{"code":525},"\nexport HUB_WHITE_LIST_PATHS=/home/work\n",[527],{"type":18,"tag":480,"props":528,"children":529},{"__ignoreMap":7},[530],{"type":24,"value":525},{"type":18,"tag":26,"props":532,"children":533},{},[534],{"type":24,"value":535},"执行以下 Python 脚本从魔乐社区下载昇思 MindSpore 版本的 Qwen3-32B 文件至指定路径/home/work。下载的文件包含模型代码、权重、分词模型和示例代码，占用约 62GB 的磁盘空间：",{"type":18,"tag":475,"props":537,"children":539},{"code":538},"\nfrom openmind_hub import snapshot_download\nsnapshot_download(\n    repo_id=\"MindSpore-Lab/Qwen3-32B\",\n    local_dir=\"/home/work\",\n    local_dir_use_symlinks=False\n)\n",[540],{"type":18,"tag":480,"props":541,"children":542},{"__ignoreMap":7},[543],{"type":24,"value":538},{"type":18,"tag":26,"props":545,"children":546},{},[547],{"type":24,"value":548},"注意事项：",{"type":18,"tag":266,"props":550,"children":551},{},[552],{"type":18,"tag":270,"props":553,"children":554},{},[555],{"type":24,"value":556},"下载时间可能因网络环境而异，建议在稳定的网络环境下操作。",{"type":18,"tag":26,"props":558,"children":559},{},[560,562],{"type":24,"value":561},"2 ",{"type":18,"tag":228,"props":563,"children":564},{},[565],{"type":24,"value":566},"服务化部署",{"type":18,"tag":26,"props":568,"children":569},{},[570,575,576],{"type":18,"tag":228,"props":571,"children":572},{},[573],{"type":24,"value":574},"1.",{"type":24,"value":234},{"type":18,"tag":228,"props":577,"children":578},{},[579],{"type":18,"tag":228,"props":580,"children":581},{},[582],{"type":24,"value":583},"添加环境变量",{"type":18,"tag":26,"props":585,"children":586},{},[587],{"type":24,"value":588},"在服务器中添加如下环境变量：",{"type":18,"tag":475,"props":590,"children":592},{"code":591},"\nexport MINDFORMERS_MODEL_CONFIG=/home/work/Qwen3-32B/predict_qwen3_32b.yaml\nexport ASCEND_CUSTOM_PATH=$ASCEND_HOME_PATH/../\nexport vLLM_MODEL_BACKEND=MindFormers\nexport vLLM_MODEL_MEMORY_USE_GB=50\nexport ASCEND_TOTAL_MEMORY_GB=64\nexport MS_ENABLE_LCCL=off\nexport HCCL_OP_EXPANSION_MODE=AIV\nexport HCCL_SOCKET_IFNAME=enp189s0f0\nexport GLOO_SOCKET_IFNAME=enp189s0f0\nexport TP_SOCKET_IFNAME=enp189s0f0\nexport HCCL_CONNECT_TIMEOUT=3600\nexport ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\n",[593],{"type":18,"tag":480,"props":594,"children":595},{"__ignoreMap":7},[596],{"type":24,"value":591},{"type":18,"tag":26,"props":598,"children":599},{},[600,605,606],{"type":18,"tag":228,"props":601,"children":602},{},[603],{"type":24,"value":604},"2.",{"type":24,"value":234},{"type":18,"tag":228,"props":607,"children":608},{},[609],{"type":18,"tag":228,"props":610,"children":611},{},[612],{"type":24,"value":613},"拉起服务",{"type":18,"tag":26,"props":615,"children":616},{},[617],{"type":24,"value":618},"执行以下命令拉起服务：",{"type":18,"tag":475,"props":620,"children":622},{"code":621},"python3 -m vllm_mindspore.entrypoints vllm.entrypoints.openai.api_server --model \"Qwen3-32B\" --trust_remote_code --tensor_parallel_size=32 --enable-prefix-caching --enable-chunked-prefill --max-num-seqs=256 --block-size=32 --max_model_len=70000 --max-num-batched-tokens=2048 --distributed-executor-backend=ray\n",[623],{"type":18,"tag":480,"props":624,"children":625},{"__ignoreMap":7},[626],{"type":24,"value":621},{"type":18,"tag":26,"props":628,"children":629},{},[630,635,636],{"type":18,"tag":228,"props":631,"children":632},{},[633],{"type":24,"value":634},"3.",{"type":24,"value":234},{"type":18,"tag":228,"props":637,"children":638},{},[639],{"type":18,"tag":228,"props":640,"children":641},{},[642],{"type":24,"value":643},"执行推理请求测试",{"type":18,"tag":26,"props":645,"children":646},{},[647],{"type":24,"value":648},"执行以下命令发送流式推理请求进行测试：",{"type":18,"tag":475,"props":650,"children":652},{"code":651},"curl http://localhost:8000/v1/completions -H \"Content-Type: application/json\" -d '{\"model\": \"Qwen3-32B \",\"prompt\": \"请介绍下北京的top景点\", \"temperature\": 0, \"max_tokens\": 256, \"top_p\": 1.0, \"top_k\": 1, \"repetition_penalty\":1.0}'\n",[653],{"type":18,"tag":480,"props":654,"children":655},{"__ignoreMap":7},[656],{"type":24,"value":651},{"type":18,"tag":26,"props":658,"children":659},{},[660],{"type":24,"value":661},"本文档提供的模型代码、权重文件和部署镜像，当前仅限于基于昇思MindSpore AI框架体验Qwen3-32B的部署效果，不支持生产环境部署。相关使用问题请反馈至Issue。",{"type":18,"tag":26,"props":663,"children":664},{},[665],{"type":24,"value":666},"昇思MindSpore AI框架将持续支持相关主流模型演进，并根据开源情况面向全体开发者提供镜像与支持。",{"title":7,"searchDepth":668,"depth":668,"links":669},4,[670,673,675],{"id":225,"depth":671,"text":672},3,"# 01 Qwen3模型介绍",{"id":364,"depth":671,"text":674},"# 02 昇思MindSpore相关技术特性",{"id":410,"depth":671,"text":676},"# 03 手把手教程：基于昇思快速上手Qwen3模型训练和推理","markdown","content:technology-blogs:zh:3709.md","content","technology-blogs/zh/3709.md","technology-blogs/zh/3709","md",1776506133561]