\u4ee3\u7801\u7ea7\u4f18\u5316\u6280\u672f<\/strong><\/h2>\n

\u667a\u80fd\u7684\u7f16\u7801\u5b9e\u8df5\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u8bad\u7ec3\u6548\u7387\u3002\u4ee5\u4e0b\u662f\u5982\u4f55\u4f18\u5316\u4ee3\u7801\u4ee5\u83b7\u5f97\u6700\u4f73\u6027\u80fd\uff1a<\/p>\n

\n
• \u6279\u91cf\u5927\u5c0f\u4f18\u5316\uff1a\n
  \n
  • \u4ece2\u7684\u5e42\u6b21\u65b9\u6279\u91cf\u5927\u5c0f\u5f00\u59cb\uff0832\u300164\u3001128\uff09<\/li>\n
  • \u4f7f\u7528\u68af\u5ea6\u7d2f\u79ef\u5b9e\u73b0\u66f4\u5927\u7684\u6709\u6548\u6279\u91cf<\/li>\n
  • \u76d1\u63a7\u5185\u5b58\u4f7f\u7528\u4e0e\u8bad\u7ec3\u7a33\u5b9a\u6027<\/li>\n<\/ul>\n<\/li>\n
  • \u5185\u5b58\u7ba1\u7406\uff1a\n
    \n
    • \u5b9e\u73b0\u68af\u5ea6\u68c0\u67e5\u70b9<\/li>\n
    • \u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff08FP16\/BF16\uff09<\/li>\n
    • \u5728\u8bad\u7ec3\u8fed\u4ee3\u4e4b\u95f4\u6e05\u9664\u7f13\u5b58<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

      \u4ee5\u4e0b\u662f\u5b9e\u73b0\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u7684\u5b9e\u7528\u793a\u4f8b\uff1a<\/p>\n

      \r\nimport torch\r\nfrom torch.cuda.amp import autocast, GradScaler\r\n\r\nscaler = GradScaler()\r\nfor data in dataloader:\r\n    with autocast():\r\n        output = model(data)\r\n        loss = criterion(output)\r\n    scaler.scale(loss).backward()\r\n    scaler.step(optimizer)\r\n    scaler.update()\r\n<\/code><\/pre>\n
      \u6570\u636e\u7ba1\u9053\u4f18\u5316<\/strong><\/h2>\n
      \u9ad8\u6548\u7684\u6570\u636e\u5904\u7406\u5bf9\u4e8e\u7ef4\u6301\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u6700\u4f73\u5229\u7528\u7387\u81f3\u5173\u91cd\u8981\u3002\u8003\u8651\u8fd9\u4e9b\u9ad8\u7ea7\u6280\u672f\uff1a<\/p>\n
      \n
      \u6570\u636e\u52a0\u8f7d\uff1a\n
      \n
      \u4f7f\u7528NVIDIA DALI\u8fdb\u884cGPU\u52a0\u901f\u6570\u636e\u52a0\u8f7d<\/li>\n
      \u5b9e\u73b0\u9884\u53d6\u673a\u5236<\/li>\n
      \u4f18\u5316\u6570\u636e\u96c6\u683c\u5f0f\uff08TFRecord\u3001WebDataset\uff09<\/li>\n<\/ul>\n<\/li>\n
      \u5b58\u50a8\u89e3\u51b3\u65b9\u6848\uff1a\n
      \n
      \u4f7f\u7528NVMe SSD\u4ee5\u83b7\u5f97\u66f4\u5feb\u7684I\/O<\/li>\n
      \u5b9e\u73b0\u6570\u636e\u5206\u7247<\/li>\n
      \u5bf9\u5c0f\u578b\u6570\u636e\u96c6\u8003\u8651\u57fa\u4e8eRAM\u7684\u6570\u636e\u96c6<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n
      \u6846\u67b6\u7279\u5b9a\u4f18\u5316<\/strong><\/h2>\n
      \u4e0d\u540c\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e3a\u7f8e\u56fdGPU\u670d\u52a1\u5668\u63d0\u4f9b\u72ec\u7279\u7684\u4f18\u5316\u673a\u4f1a\uff1a<\/p>\n
      \n
      PyTorch\u4f18\u5316\uff1a\n
      \n
      \u542f\u7528JIT\u7f16\u8bd1<\/li>\n
      \u5bf9PyTorch 2.0+\u4f7f\u7528torch.compile()<\/li>\n
      \u5b9e\u73b0DistributedDataParallel<\/li>\n<\/ul>\n<\/li>\n
      TensorFlow\u4f18\u5316\uff1a\n
      \n
      \u542f\u7528XLA\u7f16\u8bd1<\/li>\n
      \u4f7f\u7528tf.function\u88c5\u9970\u5668<\/li>\n
      \u5b9e\u73b0tf.distribute\u7b56\u7565<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u76d1\u63a7\u548c\u6027\u80fd\u8ddf\u8e2a<\/strong><\/h2>\n
      \u5b9e\u65bd\u5f3a\u5927\u7684\u76d1\u63a7\u7cfb\u7edf\u786e\u4fdd\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u6301\u7eed\u4f18\u5316\uff1a<\/p>\n
      \n
      \u5173\u952e\u6307\u6807\u8ddf\u8e2a\uff1a\n
      \n
      GPU\u4f7f\u7528\u7387\uff08\u76ee\u6807>90%\uff09<\/li>\n
      \u5185\u5b58\u4f7f\u7528\u6a21\u5f0f<\/li>\n
      PCIe\u5e26\u5bbd\u5229\u7528\u7387<\/li>\n
      \u6e29\u5ea6\u6307\u6807<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u4f7f\u7528\u8fd9\u4e2a\u7b80\u5355\u7684Python\u811a\u672c\u8fdb\u884c\u57fa\u672cGPU\u76d1\u63a7\uff1a<\/p>\n
      \r\nimport nvidia_smi\r\n\r\ndef monitor_gpu():\r\n    nvidia_smi.nvmlInit()\r\n    handle = nvidia_smi.nvmlDeviceGetHandleByIndex(0)\r\n    info = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)\r\n    util = nvidia_smi.nvmlDeviceGetUtilizationRates(handle)\r\n    print(f\"\u5185\u5b58\uff1a{info.used\/1024**2:.2f}MB\")\r\n    print(f\"\u4f7f\u7528\u7387\uff1a{util.gpu}%\")\r\n<\/code><\/pre>\n
      \u5e38\u89c1\u6027\u80fd\u95ee\u9898\u6545\u969c\u6392\u9664<\/strong><\/h2>\n
      \u89e3\u51b3\u8fd9\u4e9b\u9891\u7e41\u51fa\u73b0\u7684\u74f6\u9888\u4ee5\u7ef4\u6301\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u6700\u4f73\u8bad\u7ec3\u901f\u5ea6\uff1a<\/p>\n
      \n
      \u5185\u5b58\u95ee\u9898\uff1a\n
      \n
      \u5185\u5b58\u6ea2\u51fa\u9519\u8bef<\/li>\n
      \u5185\u5b58\u788e\u7247\u5316<\/li>\n
      \u7f13\u5b58\u6ea2\u51fa<\/li>\n<\/ul>\n<\/li>\n
      \u5904\u7406\u74f6\u9888\uff1a\n
      \n
      CPU\u74f6\u9888<\/li>\n
      I\/O\u9650\u5236<\/li>\n
      \u7f51\u7edc\u5e26\u5bbd\u9650\u5236<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n
      \u6700\u4f73\u5b9e\u8df5\u548c\u9762\u5411\u672a\u6765<\/strong><\/h2>\n
      \u901a\u8fc7\u8fd9\u4e9b\u7b56\u7565\u7ef4\u6301\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u957f\u671f\u4f18\u5316\uff1a<\/p>\n
      \n
      \u5b9a\u671f\u7ef4\u62a4\uff1a\n
      \n
      \u6bcf\u5468\u9a71\u52a8\u7a0b\u5e8f\u66f4\u65b0<\/li>\n
      \u6bcf\u6708\u6027\u80fd\u5ba1\u8ba1<\/li>\n
      \u5b63\u5ea6\u786c\u4ef6\u68c0\u67e5<\/li>\n<\/ul>\n<\/li>\n
      \u672a\u6765\u8003\u8651\uff1a\n
      \n
      \u89c4\u5212\u53ef\u6269\u5c55\u6027<\/li>\n
      \u53ca\u65f6\u4e86\u89e3\u6700\u65b0GPU\u6280\u672f<\/li>\n
      \u8003\u8651\u4e91GPU\u670d\u52a1\u5668\u79df\u7528\u66ff\u4ee3\u65b9\u6848<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u7ed3\u8bba<\/strong><\/h2>\n
      \u4f18\u5316\u7f8e\u56fdGPU\u670d\u52a1\u5668\u8bad\u7ec3\u901f\u5ea6\u9700\u8981\u6574\u4f53\u65b9\u6cd5\uff0c\u7ed3\u5408\u786c\u4ef6\u4e13\u4e1a\u77e5\u8bc6\u548c\u8f6f\u4ef6\u6280\u5de7\u3002\u901a\u8fc7\u5b9e\u65bd\u8fd9\u4e9b\u5148\u8fdb\u7684\u4f18\u5316\u6280\u672f\uff0c\u60a8\u53ef\u4ee5\u663e\u8457\u63d0\u5347GPU\u670d\u52a1\u5668\u6027\u80fd\u548c\u8bad\u7ec3\u6548\u7387\u3002\u8bf7\u8bb0\u4f4f\uff0c\u7f8e\u56fdGPU\u670d\u52a1\u5668\u4f18\u5316\u662f\u4e00\u4e2a\u9700\u8981\u5b9a\u671f\u76d1\u63a7\u548c\u66f4\u65b0\u7684\u6301\u7eed\u8fc7\u7a0b\uff0c\u4ee5\u4fdd\u6301\u6700\u4f73\u6027\u80fd\u3002<\/p>\n
      \u65e0\u8bba\u60a8\u662f\u4f7f\u7528\u7f8e\u56fdGPU\u670d\u52a1\u5668\u79df\u7528\u670d\u52a1\u8fd8\u662f\u7ba1\u7406\u81ea\u5df1\u7684\u670d\u52a1\u5668\u6258\u7ba1\u8bbe\u7f6e\uff0c\u8fd9\u4e9b\u4f18\u5316\u7b56\u7565\u90fd\u5c06\u5e2e\u52a9\u60a8\u5b9e\u73b0\u6700\u5927\u8bad\u7ec3\u901f\u5ea6\u548c\u6700\u4f73\u8d44\u6e90\u5229\u7528\u3002\u5728\u4f18\u5316\u5de5\u4f5c\u4e2d\u4fdd\u6301\u79ef\u6781\u4e3b\u52a8\uff0c\u5e76\u968f\u7740\u6280\u672f\u53d1\u5c55\u52c7\u4e8e\u5c1d\u8bd5\u65b0\u6280\u672f\u3002<\/p>\n
       \u91cd\u70b9\u5173\u952e\u8bcd\u5728US GPU Server<\/p>\n","protected":false},"excerpt":{"rendered":"
      \u5728\u5f53\u4eca\u7684\u79d1\u6280\u73af\u5883\u4e2d\uff0c\u4f18\u5316\u7f8e\u56fdGPU\u670d\u52a1\u5668\u5728AI\u8bad\u7ec3\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u53d8\u5f97\u81f3\u5173\u91cd\u8981\u3002\u65e0\u8bba\u662f\u8fd0\u884c\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u8fd8\u662f\u5904\u7406\u6d77\u91cf\u6570\u636e\u96c6\uff0c\u4f18\u5316GPU\u670d\u52a1\u5668\u7684\u8bad\u7ec3\u901f\u5ea6\u90fd\u80fd\u663e\u8457\u5f71\u54cd\u9879\u76ee\u7684\u65f6\u95f4\u7ebf\u548c\u6548\u7387\u3002\u672c\u7efc\u5408\u6307\u5357\u6df1\u5165\u63a2\u8ba8\u5df2\u9a8c\u8bc1\u7684\u4f18\u5316\u6280\u672f\uff0c\u91cd\u70b9\u5173\u6ce8\u7f8e\u56fdGPU\u670d\u52a1\u5668\u4f18\u5316\u548c\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u3002 \u786c\u4ef6\u7ea7\u4f18\u5316\u6280\u672f \u5353\u8d8a\u7684\u7f8e\u56fdGPU\u670d\u52a1\u5668\u6027\u80fd\u57fa\u7840\u5728\u4e8e\u786c\u4ef6\u914d\u7f6e\u3002\u8ba9\u6211\u4eec\u63a2\u8ba8\u80fd\u51b3\u5b9a\u8bad\u7ec3\u901f\u5ea6\u7684\u5173\u952e\u7ec4\u4ef6\uff1a GPU\u9009\u62e9\uff1a\u5728NVIDIA\u7684\u5f3a\u5927\u4ea7\u54c1\u4e2d\u9009\u62e9\uff1a A100\uff1a\u6700\u9002\u5408\u5927\u89c4\u6a21\u4f01\u4e1a\u5de5\u4f5c\u8d1f\u8f7d V100\uff1a\u51fa\u8272\u7684\u6027\u4ef7\u6bd4 H100\uff1a\u6700\u65b0\u4e00\u4ee3\u5c16\u7aef\u6027\u80fd \u591aGPU\u8bbe\u7f6e\uff1a\u914d\u7f6e\u5177\u6709\u9002\u5f53NVLink\u8fde\u63a5\u7684\u591a\u4e2aGPU PCIe\u5e26\u5bbd\uff1a\u786e\u4fddPCIe 4.0\u6216\u66f4\u65b0\u7248\u672c\u4ee5\u5b9e\u73b0\u6700\u4f73\u6570\u636e\u4f20\u8f93 \u5185\u5b58\u914d\u7f6e\uff1a\u5e73\u8861GPU\u5185\u5b58\u548c\u7cfb\u7edfRAM \u7cfb\u7edf\u7ea7\u4f18\u5316\u7b56\u7565 \u9002\u5f53\u7684\u7cfb\u7edf\u914d\u7f6e\u53ef\u4ee5\u91ca\u653e\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u6f5c\u5728\u6027\u80fd\uff1a CUDA\u73af\u5883\uff1a \u5b89\u88c5\u6700\u65b0\u7684CUDA\u5de5\u5177\u5305\uff0811.8\u6216\u66f4\u65b0\u7248\u672c\uff09 \u5b9a\u671f\u66f4\u65b0NVIDIA\u9a71\u52a8\u7a0b\u5e8f \u914d\u7f6eCUDA\u8ba1\u7b97\u80fd\u529b \u64cd\u4f5c\u7cfb\u7edf\u8c03\u4f18\uff1a \u7981\u7528\u4e0d\u5fc5\u8981\u7684\u7cfb\u7edf\u670d\u52a1 \u4f18\u5316\u5185\u6838\u53c2\u6570 \u914d\u7f6eCPU\u8c03\u901f\u5668\u4ee5\u63d0\u9ad8\u6027\u80fd \u4ee3\u7801\u7ea7\u4f18\u5316\u6280\u672f \u667a\u80fd\u7684\u7f16\u7801\u5b9e\u8df5\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u7f8e\u56fdGPU\u670d\u52a1\u5668\u7684\u8bad\u7ec3\u6548\u7387\u3002\u4ee5\u4e0b\u662f\u5982\u4f55\u4f18\u5316\u4ee3\u7801\u4ee5\u83b7\u5f97\u6700\u4f73\u6027\u80fd\uff1a \u6279\u91cf\u5927\u5c0f\u4f18\u5316\uff1a \u4ece2\u7684\u5e42\u6b21\u65b9\u6279\u91cf\u5927\u5c0f\u5f00\u59cb\uff0832\u300164\u3001128\uff09 \u4f7f\u7528\u68af\u5ea6\u7d2f\u79ef\u5b9e\u73b0\u66f4\u5927\u7684\u6709\u6548\u6279\u91cf \u76d1\u63a7\u5185\u5b58\u4f7f\u7528\u4e0e\u8bad\u7ec3\u7a33\u5b9a\u6027 \u5185\u5b58\u7ba1\u7406\uff1a \u5b9e\u73b0\u68af\u5ea6\u68c0\u67e5\u70b9 \u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff08FP16\/BF16\uff09 \u5728\u8bad\u7ec3\u8fed\u4ee3\u4e4b\u95f4\u6e05\u9664\u7f13\u5b58 \u4ee5\u4e0b\u662f\u5b9e\u73b0\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u7684\u5b9e\u7528\u793a\u4f8b\uff1a import torch from torch.cuda.amp import autocast, GradScaler scaler = GradScaler() for data in dataloader: with autocast(): output = model(data) loss = criterion(output) scaler.scale(loss).backward() […]<\/p>\n
      Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":29013,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4816],"tags":[9552,9553,9554,4281,3357],"class_list":["post-29019","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-sc","tag-us-gpu-server-optimization","tag-training-speed-improvement","tag-deep-learning-performance","tag-gpu-hosting-sc","tag-ai-acceleration-sc"],"acf":[],"yoast_head":"\n