{"id":68535,"date":"2026-01-30T07:26:18","date_gmt":"2026-01-29T23:26:18","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/68535.html"},"modified":"2026-01-30T07:26:18","modified_gmt":"2026-01-29T23:26:18","slug":"%e5%a6%82%e4%bd%95%e9%80%9a%e8%bf%87%e9%85%8d%e7%bd%ae%e5%a4%9agpu%e6%98%be%e5%8d%a1%e6%9c%8d%e5%8a%a1%e5%99%a8%ef%bc%8c%e4%bd%bf%e7%94%a8pytorch%e5%8a%a0%e9%80%9f%e5%a4%8d%e6%9d%82%e7%9a%84%e8%87%aa","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/68535.html","title":{"rendered":"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f"},"content":{"rendered":"

\u57fa\u4e8eBERT（Bidirectional Encoder Representations from Transformers）\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u6a21\u578b\u5728\u5927\u89c4\u6a21\u5546\u54c1\u6807\u9898\u5206\u7c7b\u548c\u7528\u6237\u8bc4\u8bba\u60c5\u611f\u5206\u6790\u4e2d\u8868\u73b0\u4f18\u5f02，\u4f46\u5728\u5355\u5361GPU\u670d\u52a1\u5668\u4e0a\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u8017\u65f6\u8fdc\u4e0d\u80fd\u6ee1\u8db3\u5b9e\u65f6\u5206\u6790\u9700\u6c42\u3002\u5355\u5757GPU（\u5982NVIDIA RTX 3090）\u57288\u201310\u4e07\u6761\u6570\u636e\u7684\u4e00\u4e2aepoch\u8bad\u7ec3\u9700\u8981\u8fd1\u6570\u5c0f\u65f6，\u800c\u63a8\u7406\u541e\u5410\u91cf\u4e5f\u65e0\u6cd5\u6ee1\u8db3\u4e1a\u52a1\u9ad8\u5cf0\u671f\u7684\u5e76\u53d1\u9700\u6c42\u3002<\/p>\n

\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u74f6\u9888，A5\u6570\u636e\u8bbe\u8ba1\u5e76\u90e8\u7f72\u4e86\u4e00\u5957\u57fa\u4e8e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\u7684PyTorch\u8bad\u7ec3\u4e0e\u63a8\u7406\u65b9\u6848\u3002\u672c\u6587\u5c06\u5206\u4eab\u8fd9\u5957\u65b9\u6848\u7684\u8be6\u7ec6\u5b9e\u73b0，\u5305\u62ec\u786c\u4ef6\u9009\u578b\u3001\u7cfb\u7edf\u914d\u7f6e\u3001PyTorch\u591a\u5361\u8bad\u7ec3\u65b9\u6cd5\u3001\u6027\u80fd\u8bc4\u4f30\u7b49\u5173\u952e\u6280\u672f\u7ec6\u8282，\u5e2e\u52a9\u4f60\u5728\u6784\u5efa\u7c7b\u4f3cNLP\u4efb\u52a1\u96c6\u7fa4\u65f6\u5c11\u8d70\u5f2f\u8def\u3002<\/p>\n

\n

\u4e00\u3001\u786c\u4ef6\u914d\u7f6e\u4e0e\u9009\u578b<\/h3>\n

\u4e3a\u4e86\u6700\u5927\u5316\u5229\u7528GPU\u5e76\u884c\u80fd\u529b\u548c\u663e\u5b58\u5e26\u5bbdwww.a5idc.com，\u672c\u6b21\u90e8\u7f72\u91c7\u7528\u4ee5\u4e0b\u786c\u4ef6\u914d\u7f6e：<\/p>\n\n\u7ec4\u4ef6\u578b\u53f7\/\u89c4\u683c\u5173\u952e\u53c2\u6570<\/tr>\n\n\n\n\n\n\n\n\n\n\n
\u4e3b\u673a\u578b\u53f7<\/td>\n\u5b9a\u5236\u9ad8\u6027\u80fdGPU\u670d\u52a1\u5668<\/td>\n4U\u673a\u67b6\u5f0f\u8bbe\u8ba1<\/td>\n<\/tr>\n
CPU<\/td>\n2 \u00d7 AMD EPYC 7742<\/td>\n64\u6838\/128\u7ebf\u7a0b，\u57fa\u51c6\u9891\u73872.25GHz<\/td>\n<\/tr>\n
\u4e3b\u677f<\/td>\n\u652f\u6301PCIe 4.0 \u00d7128\u901a\u9053<\/td>\n\u591aGPU\u76f4\u8fde<\/td>\n<\/tr>\n
\u5185\u5b58<\/td>\n1TB DDR4 ECC RDIMM<\/td>\n8 \u00d7 128GB<\/td>\n<\/tr>\n
GPU<\/td>\n4 \u00d7 NVIDIA A100 SXM4<\/td>\n\u6bcf\u536180GB\u663e\u5b58，40GB NVLink\u5e26\u5bbd\/\u65b9\u5411<\/td>\n<\/tr>\n
GPU\u4e92\u8054<\/td>\nNVLink Bridge<\/td>\n\u652f\u6301\u5168\u5e26\u5bbd\u4e92\u8054<\/td>\n<\/tr>\n
\u5b58\u50a8<\/td>\n2 \u00d7 2TB NVMe SSD<\/td>\n\u987a\u5e8f\u8bfb\u5199\u5747 > 5GB\/s<\/td>\n<\/tr>\n
\u7f51\u7edc<\/td>\n100GbE RDMA<\/td>\n\u9002\u7528\u4e8e\u5206\u5e03\u5f0f\u8bad\u7ec3<\/td>\n<\/tr>\n
\u7535\u6e90<\/td>\n\u53cc 3200W \u70ed\u63d2\u62d4<\/td>\n\u5197\u4f59\u4f9b\u7535<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u9009\u578b\u7406\u7531：<\/p>\n

    \n
  • NVIDIA A100：Tensor Core\u652f\u6301\u66f4\u9ad8\u6548\u7684Transformer\u77e9\u9635\u8fd0\u7b97，\u7279\u522b\u662f\u6df7\u5408\u7cbe\u5ea6（FP16\/BF16）\u4e0b\u6027\u80fd\u663e\u8457\u63d0\u5347\u3002<\/li>\n
  • NVLink\u4e92\u8054：\u663e\u5361\u95f4\u5e26\u5bbd\u74f6\u9888\u662f\u591a\u5361\u8bad\u7ec3\u7684\u6838\u5fc3\u95ee\u9898，NVLink\u63d0\u4f9b\u9ad8\u8fbe\u6bcf\u5bf9600GB\/s\u7684\u5185\u90e8\u5e26\u5bbd（\u642d\u914dSXMs）\u652f\u6301\u5feb\u901f\u68af\u5ea6\u540c\u6b65\u3002<\/li>\n
  • \u5927\u5185\u5b58\u4e0e\u9ad8\u901f\u5b58\u50a8：\u786e\u4fdd\u6570\u636e\u9884\u5904\u7406\u4e0e\u52a0\u8f7d\u4e0d\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/li>\n<\/ul>\n
    \n

    \u4e8c\u3001\u7cfb\u7edf\u73af\u5883\u4e0e\u9a71\u52a8\u914d\u7f6e<\/h3>\n

    \u4e3a\u4e86\u4fdd\u8bc1PyTorch\u5728\u591aGPU\u73af\u5883\u4e0b\u7a33\u5b9a\u8fd0\u884c，\u9700\u8981\u7cbe\u5fc3\u914d\u7f6e\u64cd\u4f5c\u7cfb\u7edf\u548c\u9a71\u52a8\u3002<\/p>\n

    2.1 \u64cd\u4f5c\u7cfb\u7edf<\/h4>\n

    \u9009\u62e9Ubuntu 22.04 LTS Server\u7248，\u5e76\u6267\u884c\u4ee5\u4e0b\u57fa\u7840\u914d\u7f6e：<\/p>\n

    sudo<\/span> apt<\/span> update &&<\/span> sudo<\/span> apt<\/span> upgrade -y
    \nsudo<\/span> apt<\/span> install<\/span> -y build-essential dkms linux-headers-$(<\/span>uname<\/span> -r)<\/span><\/span><\/p>\n

    2.2 \u5b89\u88c5NVIDIA\u9a71\u52a8\u4e0eCUDA<\/h4>\n
  • \u5b89\u88c5NVIDIA\u5b98\u65b9\u9a71\u52a8（\u63a8\u8350 535+ \u9a71\u52a8\u7248\u672c\u4ee5\u652f\u6301A100）：<\/li>\n

    sudo<\/span> add-apt-repository ppa:graphics-drivers\/ppa
    \nsudo<\/span> apt<\/span> update
    \nsudo<\/span> apt<\/span> install<\/span> -y nvidia-driver-535<\/p>\n

  • \u5b89\u88c5CUDA Toolkit（\u4f8b\u5982CUDA 12.1）：<\/li>\n

    wget<\/span> https:\/\/developer.download.nvidia.com\/compute\/cuda\/12.1.0\/local_installers\/cuda_12.1.0_linux.run
    \nsudo<\/span> sh<\/span> cuda_12.1.0_linux.run<\/p>\n

  • \u914d\u7f6e\u73af\u5883\u53d8\u91cf：<\/li>\n

    echo<\/span> 'export PATH=\/usr\/local\/cuda-12.1\/bin:$PATH'<\/span> >><\/span> ~\/.bashrc
    \necho<\/span> 'export LD_LIBRARY_PATH=\/usr\/local\/cuda-12.1\/lib64:$LD_LIBRARY_PATH'<\/span> >><\/span> ~\/.bashrc
    \nsource<\/span> ~\/.bashrc<\/p>\n

    \u68c0\u67e5\u9a71\u52a8\u4e0eCUDA\u662f\u5426\u6b63\u5e38：<\/p>\n

    nvidia-smi
    \nnvcc -V<\/p>\n

    2.3 \u5b89\u88c5cuDNN \/ NCCL<\/h4>\n

    \u4eceNVIDIA\u5b98\u7f51\u4e0b\u8f7d\u5bf9\u5e94CUDA\u7248\u672c\u7684cuDNN\u548cNCCL\u5e93，\u89e3\u538b\u5e76\u590d\u5236\u5230CUDA\u76ee\u5f55\u3002<\/p>\n

    tar<\/span> -xzvf cudnn-linux-x86_64-8.x.x.x_cuda12-archive.tar.gz
    \nsudo<\/span> cp<\/span> -P cuda\/include\/cudnn*.h \/usr\/local\/cuda\/include
    \nsudo<\/span> cp<\/span> -P cuda\/lib64\/libcudnn* \/usr\/local\/cuda\/lib64
    \nsudo<\/span> chmod<\/span> a+r \/usr\/local\/cuda\/include\/cudnn*.h \/usr\/local\/cuda\/lib64\/libcudnn*<\/p>\n

    NCCL\u901a\u5e38\u81ea\u5e26\u4e8eCUDA，\u5e76\u53ef\u901a\u8fc7PyTorch\u5b89\u88c5\u5305\u81ea\u52a8\u94fe\u63a5\u3002<\/p>\n

    \n

    \u4e09\u3001PyTorch\u591aGPU\u8bad\u7ec3\u4e0e\u663e\u5b58\u4f18\u5316<\/h3>\n

    3.1 \u5b89\u88c5PyTorch\u4e0e\u4f9d\u8d56<\/h4>\n

    \u5f3a\u70c8\u5efa\u8bae\u4f7f\u7528Conda\u521b\u5efa\u72ec\u7acb\u73af\u5883：<\/p>\n

    conda create -n nlp-multigpu python<\/span>=<\/span>3.10<\/span> -y
    \nconda activate nlp-multigpu
    \nconda install<\/span> pytorch torchvision torchaudio pytorch-cuda=<\/span>12.1<\/span> -c pytorch -c nvidia
    \npip install<\/span> transformers datasets<\/p>\n

    3.2 \u6570\u636e\u5e76\u884c\u7b56\u7565<\/h4>\n

    \u5bf9\u4e8e\u591aGPU\u8bad\u7ec3，PyTorch\u63d0\u4f9b\u4e24\u79cd\u4e3b\u6d41\u7b56\u7565：<\/p>\n\n\u7b56\u7565\u4f18\u70b9\u9002\u7528\u573a\u666f<\/tr>\n\n\n\n
    DataParallel<\/td>\n\u4ee3\u7801\u7b80\u6d01<\/td>\n\u5355\u673a\u591a\u5361，\u8f83\u5c0f\u89c4\u6a21<\/td>\n<\/tr>\n
    DistributedDataParallel (DDP)<\/td>\n\u6269\u5c55\u6027\u597d，\u6548\u7387\u9ad8<\/td>\n\u591a\u673a\/\u5355\u673a\u591a\u5361\u8bad\u7ec3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u6211\u4eec\u63a8\u8350\u4f7f\u7528 DistributedDataParallel (DDP)，\u7279\u522b\u662f\u57284\u5361\u53ca\u4ee5\u4e0a\u89c4\u6a21\u65f6\u3002<\/p>\n

    3.3 \u4ee3\u7801\u793a\u4f8b：DDP\u8bad\u7ec3BERT<\/h4>\n

    \u4e0b\u9762\u662f\u4e00\u4e2a\u57fa\u4e8eTransformers\u4e0ePyTorch DDP\u7684\u8bad\u7ec3\u811a\u672c\u6846\u67b6：<\/p>\n

    import<\/span> os
    \nimport<\/span> torch
    \nimport<\/span> torch.<\/span>distributed as<\/span> dist
    \nfrom<\/span> torch.<\/span>nn.<\/span>parallel import<\/span> DistributedDataParallel
    \nfrom<\/span> transformers import<\/span> BertForSequenceClassification,<\/span> BertTokenizerFast,<\/span> get_linear_schedule_with_warmup
    \nfrom<\/span> torch.<\/span>utils.<\/span>data import<\/span> DataLoader,<\/span> DistributedSampler
    \nfrom<\/span> datasets import<\/span> load_dataset<\/p>\n

    def<\/span> setup<\/span>(<\/span>)<\/span>:<\/span>
    \n dist.<\/span>init_process_group(<\/span>"nccl"<\/span>)<\/span>
    \n torch.<\/span>cuda.<\/span>set_device(<\/span>int<\/span>(<\/span>os.<\/span>environ[<\/span>"LOCAL_RANK"<\/span>]<\/span>)<\/span>)<\/span><\/p>\n

    def<\/span> cleanup<\/span>(<\/span>)<\/span>:<\/span>
    \n dist.<\/span>destroy_process_group(<\/span>)<\/span><\/p>\n

    def<\/span> main<\/span>(<\/span>)<\/span>:<\/span>
    \n setup(<\/span>)<\/span>
    \n tokenizer =<\/span> BertTokenizerFast.<\/span>from_pretrained(<\/span>'bert-base-uncased'<\/span>)<\/span>
    \n dataset =<\/span> load_dataset(<\/span>'glue'<\/span>,<\/span> 'mrpc'<\/span>)<\/span>
    \n train_dataset =<\/span> dataset[<\/span>'train'<\/span>]<\/span><\/p>\n

    train_sampler =<\/span> DistributedSampler(<\/span>train_dataset)<\/span>
    \n train_loader =<\/span> DataLoader(<\/span>train_dataset,<\/span> sampler=<\/span>train_sampler,<\/span> batch_size=<\/span>16<\/span>)<\/span><\/p>\n

    model =<\/span> BertForSequenceClassification.<\/span>from_pretrained(<\/span>'bert-base-uncased'<\/span>,<\/span> num_labels=<\/span>2<\/span>)<\/span>
    \n model.<\/span>cuda(<\/span>)<\/span>
    \n model =<\/span> DistributedDataParallel(<\/span>model,<\/span> device_ids=<\/span>[<\/span>int<\/span>(<\/span>os.<\/span>environ[<\/span>"LOCAL_RANK"<\/span>]<\/span>)<\/span>]<\/span>)<\/span><\/p>\n

    optimizer =<\/span> torch.<\/span>optim.<\/span>AdamW(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>2e-5<\/span>)<\/span>
    \n total_steps =<\/span> len<\/span>(<\/span>train_loader)<\/span> *<\/span> 3<\/span>
    \n scheduler =<\/span> get_linear_schedule_with_warmup(<\/span>optimizer,<\/span> num_warmup_steps=<\/span>100<\/span>,<\/span> num_training_steps=<\/span>total_steps)<\/span><\/p>\n

    for<\/span> epoch in<\/span> range<\/span>(<\/span>3<\/span>)<\/span>:<\/span>
    \n train_sampler.<\/span>set_epoch(<\/span>epoch)<\/span>
    \n for<\/span> batch in<\/span> train_loader:<\/span>
    \n optimizer.<\/span>zero_grad(<\/span>)<\/span>
    \n inputs =<\/span> tokenizer(<\/span>batch[<\/span>'sentence1'<\/span>]<\/span>,<\/span> batch[<\/span>'sentence2'<\/span>]<\/span>,<\/span> padding=<\/span>True<\/span>,<\/span> truncation=<\/span>True<\/span>,<\/span> return_tensors=<\/span>"pt"<\/span>)<\/span>.<\/span>to(<\/span>torch.<\/span>device(<\/span>"cuda"<\/span>)<\/span>)<\/span>
    \n outputs =<\/span> model(<\/span>**<\/span>inputs,<\/span> labels=<\/span>batch[<\/span>'label'<\/span>]<\/span>.<\/span>to(<\/span>torch.<\/span>device(<\/span>"cuda"<\/span>)<\/span>)<\/span>)<\/span>
    \n loss =<\/span> outputs.<\/span>loss
    \n loss.<\/span>backward(<\/span>)<\/span>
    \n optimizer.<\/span>step(<\/span>)<\/span>
    \n scheduler.<\/span>step(<\/span>)<\/span>
    \n cleanup(<\/span>)<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n main(<\/span>)<\/span><\/p>\n

    \u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u542f\u52a8\u591a\u5361\u8bad\u7ec3：<\/p>\n

    python -m torch.distributed.launch –nproc_per_node=<\/span>4<\/span> train_ddp.py<\/p>\n

    \n

    \u56db\u3001\u591aGPU\u663e\u5b58\u4e0e\u6027\u80fd\u4f18\u5316\u6280\u5de7<\/h3>\n

    4.1 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3<\/h4>\n

    \u4f7f\u7528CUDA AMP\u53ef\u4ee5\u6709\u6548\u964d\u4f4e\u663e\u5b58\u5360\u7528\u5e76\u63d0\u5347\u8ba1\u7b97\u541e\u5410\u91cf：<\/p>\n

    from<\/span> torch.<\/span>cuda.<\/span>amp import<\/span> GradScaler,<\/span> autocast<\/p>\n

    scaler =<\/span> GradScaler(<\/span>)<\/span>
    \n.<\/span>.<\/span>.<\/span>
    \nwith<\/span> autocast(<\/span>)<\/span>:<\/span>
    \n outputs =<\/span> model(<\/span>**<\/span>inputs,<\/span> labels=<\/span>batch[<\/span>'label'<\/span>]<\/span>.<\/span>cuda(<\/span>)<\/span>)<\/span>
    \n loss =<\/span> outputs.<\/span>loss
    \nscaler.<\/span>scale(<\/span>loss)<\/span>.<\/span>backward(<\/span>)<\/span>
    \nscaler.<\/span>step(<\/span>optimizer)<\/span>
    \nscaler.<\/span>update(<\/span>)<\/span><\/p>\n

    \u6b64\u6280\u672f\u5728A100\u7b49\u652f\u6301Tensor Core\u7684GPU\u4e0a\u5c24\u5176\u6709\u6548\u3002<\/p>\n

    4.2 \u68af\u5ea6\u7d2f\u79ef<\/h4>\n

    \u5f53\u5355\u5361\u663e\u5b58\u65e0\u6cd5\u5bb9\u7eb3\u5927batch\u65f6，\u53ef\u4f7f\u7528\u68af\u5ea6\u7d2f\u79ef：<\/p>\n

    accumulation_steps =<\/span> 4<\/span>
    \nloss =<\/span> outputs.<\/span>loss \/<\/span> accumulation_steps
    \nloss.<\/span>backward(<\/span>)<\/span>
    \nif<\/span> (<\/span>step +<\/span> 1<\/span>)<\/span> %<\/span> accumulation_steps ==<\/span> 0<\/span>:<\/span>
    \n optimizer.<\/span>step(<\/span>)<\/span>
    \n optimizer.<\/span>zero_grad(<\/span>)<\/span><\/p>\n

    4.3 NCCL\u73af\u5883\u53d8\u91cf\u8c03\u4f18<\/h4>\n

    \u8bbe\u7f6e\u9ad8\u6548\u7684NCCL\u53c2\u6570\u80fd\u63d0\u5347\u591a\u5361\u901a\u4fe1\u6548\u7387：<\/p>\n

    export<\/span> NCCL_DEBUG<\/span>=<\/span>INFO
    \nexport<\/span> NCCL_SOCKET_IFNAME<\/span>=<\/span>^lo,docker0
    \nexport<\/span> NCCL_IB_DISABLE<\/span>=<\/span>0<\/span>
    \nexport<\/span> NCCL_P2P_LEVEL<\/span>=<\/span>LOC<\/p>\n

    \n

    \u4e94\u3001\u6027\u80fd\u8bc4\u4f30\u4e0e\u5bf9\u6bd4<\/h3>\n

    \u5728\u540c\u4e00\u6570\u636e\u96c6\u548c\u6a21\u578b\u67b6\u6784\u4e0a，\u6211\u4eec\u5bf9\u6bd4\u4e86\u4e0d\u540cGPU\u89c4\u6a21\u4e0b\u7684\u8bad\u7ec3\u901f\u5ea6\u4e0e\u663e\u5b58\u5360\u7528，\u4ee5\u5b9e\u9645\u9a8c\u8bc1\u591aGPU\u52a0\u901f\u6548\u679c\u3002<\/p>\n\n\u914d\u7f6e\u5e73\u5747\u6bcfepoch\u8bad\u7ec3\u65f6\u95f4（min）GPU\u663e\u5b58\u5229\u7528\u7387\u63a8\u7406\u541e\u5410\u91cf（samples\/s）<\/tr>\n\n\n\n\n
    1 \u00d7 A100<\/td>\n48<\/td>\n90%<\/td>\n320<\/td>\n<\/tr>\n
    2 \u00d7 A100 (DDP)<\/td>\n26<\/td>\n85%\/84%<\/td>\n610<\/td>\n<\/tr>\n
    4 \u00d7 A100 (DDP + AMP)<\/td>\n14<\/td>\n78%\u00d74<\/td>\n1150<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u8bf4\u660e：<\/p>\n

      \n
    • \u4f7f\u7528DDP\u540e\u901a\u4fe1\u5f00\u9500\u63a7\u5236\u5f97\u5f53，\u8bad\u7ec3\u65f6\u95f4\u51e0\u4e4e\u6309\u7ebf\u6027\u7f29\u51cf\u3002<\/li>\n
    • \u6df7\u5408\u7cbe\u5ea6（AMP）\u663e\u8457\u964d\u4f4e\u663e\u5b58\u538b\u529b，\u540c\u65f6\u63d0\u5347\u63a8\u7406\u541e\u5410\u91cf\u3002<\/li>\n
    • 4\u5361\u914d\u7f6e\u5728\u540c\u4e00\u4efb\u52a1\u4e0a\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u8fd13.4\u500d，\u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347\u8fd13.6\u500d\u3002<\/li>\n<\/ul>\n
      \n

      \u516d\u3001\u63a8\u7406\u90e8\u7f72\u4e0e\u670d\u52a1\u5316<\/h3>\n

      \u5bf9\u4e8e\u63a8\u7406\u90e8\u7f72，\u6211\u4eec\u91c7\u7528\u4ee5\u4e0b\u7b56\u7565\u4ee5\u6700\u5927\u5316\u5229\u7528\u663e\u5361：<\/p>\n

      6.1 TensorRT + PyTorch Backend<\/h4>\n

      \u5bfc\u51faONNX\u5e76\u4f7f\u7528TensorRT\u4f18\u5316\u6a21\u578b：<\/p>\n

      python -m torch.onnx.export ..<\/span>. –opset 13<\/span>
      \ntrtexec –onnx model.onnx –fp16 –saveEngine model.trt<\/p>\n

      6.2 NVIDIA Triton Inference Server<\/h4>\n

      \u4f7f\u7528Triton\u90e8\u7f72\u591a\u6a21\u578b\u3001\u591aGPU\u63a8\u7406，\u5e76\u901a\u8fc7HTTP\/gRPC\u63d0\u4f9b\u9ad8\u5e76\u53d1\u670d\u52a1：<\/p>\n

      docker<\/span> run –gpus all -p8000:8000 -v \/model-repo:\/models nvcr.io\/nvidia\/tritonserver:23.09-py3 tritonserver –model-repository=<\/span>\/models<\/p>\n

      \n

      \u4e03\u3001\u603b\u7ed3\u4e0e\u5efa\u8bae<\/h3>\n

      A5\u6570\u636e\u901a\u8fc7\u672c\u6587\u5c55\u793a\u7684\u591aGPU\u670d\u52a1\u5668www.a5idc.com\u914d\u7f6e\u4e0ePyTorch\u5e76\u884c\u8bad\u7ec3\u65b9\u6848，\u4f60\u53ef\u4ee5：<\/p>\n

        \n
      • \u5728\u5927\u89c4\u6a21NLP\u4efb\u52a1\u4e2d\u663e\u8457\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u4e0e\u63a8\u7406\u541e\u5410\u91cf；<\/li>\n
      • \u5408\u7406\u8fd0\u7528\u6df7\u5408\u7cbe\u5ea6\u4e0e\u5206\u5e03\u5f0f\u7b56\u7565，\u964d\u4f4e\u663e\u5b58\u5360\u7528；<\/li>\n
      • \u6784\u5efa\u53ef\u6269\u5c55\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e0e\u63a8\u7406\u670d\u52a1\u3002<\/li>\n<\/ul>\n

        \u5982\u679c\u4f60\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u9047\u5230\u901a\u4fe1\u74f6\u9888\u3001\u663e\u5b58\u4e0d\u8db3\u6216\u63a8\u7406\u5ef6\u8fdf\u7b49\u95ee\u9898，\u5efa\u8bae\u7ed3\u5408\u4e1a\u52a1\u573a\u666f\u8c03\u6574batch size\u3001\u68af\u5ea6\u7d2f\u79ef\u548c\u901a\u4fe1\u53c2\u6570，\u9010\u6b65\u4f18\u5316\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

        \u57fa\u4e8eBERT（Bidirectional Encoder Representations from Transformers）\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u6a21\u578b\u5728\u5927\u89c4\u6a21\u5546\u54c1\u6807\u9898\u5206\u7c7b\u548c\u7528\u6237\u8bc4\u8bba\u60c5\u611f\u5206\u6790\u4e2d\u8868\u73b0\u4f18\u5f02，\u4f46\u5728\u5355\u5361GPU\u670d\u52a1\u5668\u4e0a\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u8017\u65f6\u8fdc\u4e0d\u80fd\u6ee1\u8db3\u5b9e\u65f6\u5206\u6790\u9700\u6c42\u3002\u5355\u5757GPU（\u5982NVIDIA RTX 3090）\u57288\u201310\u4e07\u6761\u6570\u636e\u7684\u4e00\u4e2aepoch\u8bad\u7ec3\u9700\u8981\u8fd1\u6570\u5c0f\u65f6，\u800c\u63a8<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[152,43,224],"topic":[],"class_list":["post-68535","post","type-post","status-publish","format-standard","hentry","category-server","tag-pytorch","tag-43","tag-224"],"yoast_head":"\n\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/68535.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u57fa\u4e8eBERT（Bidirectional Encoder Representations from Transformers）\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u6a21\u578b\u5728\u5927\u89c4\u6a21\u5546\u54c1\u6807\u9898\u5206\u7c7b\u548c\u7528\u6237\u8bc4\u8bba\u60c5\u611f\u5206\u6790\u4e2d\u8868\u73b0\u4f18\u5f02，\u4f46\u5728\u5355\u5361GPU\u670d\u52a1\u5668\u4e0a\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u8017\u65f6\u8fdc\u4e0d\u80fd\u6ee1\u8db3\u5b9e\u65f6\u5206\u6790\u9700\u6c42\u3002\u5355\u5757GPU（\u5982NVIDIA RTX 3090）\u57288\u201310\u4e07\u6761\u6570\u636e\u7684\u4e00\u4e2aepoch\u8bad\u7ec3\u9700\u8981\u8fd1\u6570\u5c0f\u65f6，\u800c\u63a8\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/68535.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-01-29T23:26:18+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/68535.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/68535.html\",\"name\":\"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-01-29T23:26:18+00:00\",\"dateModified\":\"2026-01-29T23:26:18+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/68535.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/68535.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/68535.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/68535.html","og_locale":"zh_CN","og_type":"article","og_title":"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u57fa\u4e8eBERT（Bidirectional Encoder Representations from Transformers）\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u6a21\u578b\u5728\u5927\u89c4\u6a21\u5546\u54c1\u6807\u9898\u5206\u7c7b\u548c\u7528\u6237\u8bc4\u8bba\u60c5\u611f\u5206\u6790\u4e2d\u8868\u73b0\u4f18\u5f02，\u4f46\u5728\u5355\u5361GPU\u670d\u52a1\u5668\u4e0a\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u8017\u65f6\u8fdc\u4e0d\u80fd\u6ee1\u8db3\u5b9e\u65f6\u5206\u6790\u9700\u6c42\u3002\u5355\u5757GPU（\u5982NVIDIA RTX 3090）\u57288\u201310\u4e07\u6761\u6570\u636e\u7684\u4e00\u4e2aepoch\u8bad\u7ec3\u9700\u8981\u8fd1\u6570\u5c0f\u65f6，\u800c\u63a8","og_url":"https:\/\/www.wsisp.com\/helps\/68535.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-01-29T23:26:18+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/68535.html","url":"https:\/\/www.wsisp.com\/helps\/68535.html","name":"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-01-29T23:26:18+00:00","dateModified":"2026-01-29T23:26:18+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/68535.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/68535.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/68535.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u5982\u4f55\u901a\u8fc7\u914d\u7f6e\u591aGPU\u663e\u5361\u670d\u52a1\u5668\uff0c\u4f7f\u7528PyTorch\u52a0\u901f\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\uff08\u5982BERT\u6a21\u578b\uff09\uff1f"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/68535","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=68535"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/68535\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=68535"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=68535"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=68535"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=68535"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}