{"id":209326,"date":"2025-09-30T08:22:32","date_gmt":"2025-09-30T00:22:32","guid":{"rendered":"https:\/\/www.server.hk\/blog\/209326\/"},"modified":"2025-09-30T08:22:33","modified_gmt":"2025-09-30T00:22:33","slug":"vps-bert-3","status":"publish","type":"post","link":"https:\/\/server.hk\/cnblog\/209326\/","title":{"rendered":"\u9999\u6e2fVPS\u90e8\u7f72BERT\u5168\u6d41\u7a0b\u5b9e\u6218\uff1a\u5feb\u901f\u4e0a\u624b\u4e0e\u6027\u80fd\u4f18\u5316"},"content":{"rendered":"<p>\u5728\u8fb9\u7f18\u8ba1\u7b97\u548c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u65e5\u76ca\u666e\u53ca\u7684\u4eca\u5929\uff0c\u90e8\u7f72BERT\u7c7b\u6a21\u578b\u5230\u4e91\u7aef\u6210\u4e3a\u8bb8\u591a\u7f51\u7ad9\u3001\u5e94\u7528\u4ee5\u53ca\u4f01\u4e1a\u7ea7\u670d\u52a1\u7684\u5e38\u89c1\u9700\u6c42\u3002\u672c\u6587\u9762\u5411\u7ad9\u957f\u3001\u4f01\u4e1a\u7528\u6237\u548c\u5f00\u53d1\u8005\uff0c\u4ece\u539f\u7406\u3001\u90e8\u7f72\u6d41\u7a0b\u3001\u6027\u80fd\u4f18\u5316\u548c\u9009\u8d2d\u5efa\u8bae\u56db\u4e2a\u5c42\u9762\uff0c\u8bb2\u89e3\u5982\u4f55\u5728\u9999\u6e2fVPS\u4e0a\u5b8c\u6210BERT\u7684\u5168\u6d41\u7a0b\u5b9e\u6218\uff0c\u5e2e\u52a9\u4f60\u5728\u5b9e\u9645\u751f\u4ea7\u73af\u5883\u4e2d\u83b7\u5f97\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u541e\u5410\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<h2>\u5f15\u8a00\uff1a\u4e3a\u4ec0\u4e48\u9009\u62e9\u9999\u6e2fVPS\u90e8\u7f72BERT<\/h2>\n<p>BERT\u8fd9\u7c7b\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u6587\u672c\u5206\u7c7b\u3001\u95ee\u7b54\u3001\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\u7b49\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u4e5f\u5bf9\u8ba1\u7b97\u548c\u5185\u5b58\u6709\u8f83\u9ad8\u8981\u6c42\u3002<strong>\u9999\u6e2fVPS<\/strong>\u5728\u5730\u7406\u4f4d\u7f6e\u4e0a\u5bf9\u5927\u4e2d\u534e\u533a\u7528\u6237\u5177\u6709\u5929\u7136\u7684\u4f4e\u5ef6\u8fdf\u4f18\u52bf\uff0c\u540c\u65f6\u63d0\u4f9b\u826f\u597d\u7684\u5e26\u5bbd\u548c\u7f51\u7edc\u7a33\u5b9a\u6027\u3002\u4e0e\u4e4b\u76f8\u6bd4\uff0c\u9009\u62e9<span>\u7f8e\u56fdVPS<\/span>\u6216<span>\u7f8e\u56fd\u670d\u52a1\u5668<\/span>\u901a\u5e38\u5728\u4ef7\u683c\u6216\u8de8\u5927\u9646\u8bbf\u95ee\u4e0a\u6709\u4f18\u52bf\uff0c\u4f46\u53ef\u80fd\u5728\u5b9e\u65f6\u4ea4\u4e92\u573a\u666f\u51fa\u73b0\u66f4\u9ad8\u7684\u7f51\u7edc\u5ef6\u8fdf\u3002\u57fa\u4e8e\u8fd9\u79cd\u6743\u8861\uff0c\u9488\u5bf9\u9762\u5411\u4e9a\u592a\u7528\u6237\u7684\u5e94\u7528\uff0c\u90e8\u7f72\u5728\u9999\u6e2fVPS\u5f80\u5f80\u80fd\u83b7\u5f97\u66f4\u597d\u4f53\u9a8c\u3002<\/p>\n<h2>\u539f\u7406\u4e0e\u90e8\u7f72\u524d\u51c6\u5907<\/h2>\n<h3>\u7406\u89e3BERT\u63a8\u7406\u74f6\u9888<\/h3>\n<p>BERT\u7684\u63a8\u7406\u4e3b\u8981\u53d7\u4e09\u7c7b\u8d44\u6e90\u9650\u5236\uff1aCPU\/GPU\u8ba1\u7b97\u80fd\u529b\u3001\u5185\u5b58\uff08\u5c24\u5176\u662f\u6a21\u578b\u53c2\u6570\u5185\u5b58\uff09\u548c\u7f51\u7edc\u5e26\u5bbd\/\u5ef6\u8fdf\u3002\u5c0f\u6a21\u578b\uff08\u5982DistilBERT\u3001TinyBERT\uff09\u9002\u5408CPU\u90e8\u7f72\uff0c\u800c\u5927\u578bBERT-base\/large\u66f4\u9002\u5408GPU\u6216\u7ecf\u8fc7\u91cf\u5316\/\u526a\u679d\u7684\u52a0\u901f\u65b9\u6848\u3002<\/p>\n<h3>\u73af\u5883\u4e0e\u8f6f\u4ef6\u6808\u5efa\u8bae<\/h3>\n<ul>\n<li>\u64cd\u4f5c\u7cfb\u7edf\uff1a\u5efa\u8bae\u4f7f\u7528Ubuntu 20.04\u6216\u66f4\u9ad8\u7248\u672c\uff0c\u7a33\u5b9a\u4e14\u517c\u5bb9\u5927\u91cf\u6df1\u5ea6\u5b66\u4e60\u5e93\u3002<\/li>\n<li>\u57fa\u7840\u5de5\u5177\uff1a\u5b89\u88c5Python 3.8+\u3001pip\u3001virtualenv\u3002\u521b\u5efa\u72ec\u7acb\u865a\u62df\u73af\u5883\u4ee5\u907f\u514d\u4f9d\u8d56\u51b2\u7a81\u3002<\/li>\n<li>\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff1aPyTorch\u6216TensorFlow\uff0c\u63a8\u8350\u4f7f\u7528PyTorch + Hugging Face Transformers\u4ee5\u4fbf\u5feb\u901f\u4e0a\u624b\u3002<\/li>\n<li>\u52a0\u901f\u5e93\uff1a\u5728CPU\u4e0a\u542f\u7528MKL\/oneAPI\u6216OpenBLAS\uff1b\u5728GPU\u4e0a\u4f7f\u7528CUDA\u3001cuDNN\uff0c\u5e76\u8003\u8651TensorRT\u7528\u4e8e\u63a8\u7406\u4f18\u5316\u3002<\/li>\n<li>\u5bb9\u5668\u5316\uff1a\u4f7f\u7528Docker\u53ef\u4ee5\u7b80\u5316\u90e8\u7f72\u4e0e\u8fc1\u79fb\uff0c\u4fbf\u4e8e\u5728\u9999\u6e2f\u670d\u52a1\u5668\u4e0e\u7f8e\u56fdVPS\u95f4\u5207\u6362\u3002<\/li>\n<\/ul>\n<h2>\u5728\u9999\u6e2fVPS\u4e0a\u90e8\u7f72\u7684\u5b9e\u6218\u6d41\u7a0b<\/h2>\n<h3>1. \u9009\u62e9\u5408\u9002\u7684\u5b9e\u4f8b\u914d\u7f6e<\/h3>\n<p>\u90e8\u7f72\u524d\u5148\u660e\u786e\u4e1a\u52a1\u9700\u6c42\uff1aQPS\uff08\u6bcf\u79d2\u8bf7\u6c42\u6570\uff09\u3001\u6700\u5927\u5e76\u53d1\u6570\u3001\u5e73\u5747\u5ef6\u8fdf\u76ee\u6807\u3002\u4f8b\u5982\uff0c\u95ee\u7b54\u7cfb\u7edf\u8981\u6c42\u4f4e\u5ef6\u8fdf\uff08&lt;200ms\uff09\uff0c\u800c\u6279\u91cf\u6587\u672c\u5206\u7c7b\u53ef\u5bb9\u5fcd\u66f4\u9ad8\u5ef6\u8fdf\u3002\u4e3aBERT\u63a8\u7406\u63a8\u8350\u7684\u914d\u7f6e\uff1a<\/p>\n<ul>\n<li>CPU\u90e8\u7f72\uff08\u8f7b\u91cf\/\u4e2d\u7b49\u8d1f\u8f7d\uff09\uff1a4-16\u6838CPU\u300116-64GB\u5185\u5b58\u3001NVMe SSD\u3002<\/li>\n<li>GPU\u90e8\u7f72\uff08\u9ad8\u5e76\u53d1\u6216\u5927\u578b\u6a21\u578b\uff09\uff1a\u542bNVIDIA T4\/RTX\u7cfb\u5217\u7684\u5b9e\u4f8b\uff0c\u81f3\u5c1116GB GPU\u663e\u5b58\u4ee5\u653e\u4e0b\u6a21\u578b\u548c\u4e2d\u95f4\u7f13\u5b58\u3002<\/li>\n<li>\u7f51\u7edc\uff1a\u5e26\u5bbd\u548c\u9aa8\u5e72\u76f4\u8fde\u8d28\u91cf\u51b3\u5b9a\u8de8\u5883\u8bbf\u95ee\u4f53\u9a8c\uff0c\u9999\u6e2fVPS\u901a\u5e38\u5728\u8bbf\u95ee\u5927\u9646\u548c\u4e1c\u5357\u4e9a\u65f6\u5ef6\u8f83\u4f18\u3002<\/li>\n<\/ul>\n<h3>2. \u642d\u5efa\u8fd0\u884c\u73af\u5883\u4e0e\u62c9\u53d6\u6a21\u578b<\/h3>\n<p>\u5728VPS\u4e0a\u521b\u5efaPython\u865a\u62df\u73af\u5883\u5e76\u5b89\u88c5\u4f9d\u8d56\uff1apip install torch transformers flask gunicorn\u3002\u4f7f\u7528Hugging Face\u7684\u6a21\u578b\u4ed3\u5e93\uff0c\u793a\u4f8b\uff1a\u4ecetransformers\u52a0\u8f7d\u6a21\u578b\u2014model = AutoModelForSequenceClassification.from_pretrained(&#8216;bert-base-uncased&#8217;)\u3002\u82e5\u5e26\u5bbd\u6216\u9690\u79c1\u9650\u5236\uff0c\u53ef\u5148\u5728\u672c\u5730\u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6\u518d\u4e0a\u4f20\u5230VPS\u3002<\/p>\n<h3>3. \u6a21\u578b\u4f18\u5316\u4e0e\u5bfc\u51fa<\/h3>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u541e\u5410\u4e0e\u964d\u4f4e\u5ef6\u8fdf\uff0c\u5e38\u89c1\u4f18\u5316\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u6a21\u578b\u91cf\u5316\uff1a\u4f7f\u7528\u52a8\u6001\u91cf\u5316\uff08PyTorch\u7684torch.quantization.quantize_dynamic\uff09\u53ef\u5c06\u6a21\u578b\u5c3a\u5bf8\u548c\u5185\u5b58\u5360\u7528\u663e\u8457\u4e0b\u964d\uff0c\u9002\u5408CPU\u63a8\u7406\u3002<\/li>\n<li>\u56fe\u4f18\u5316\u4e0e\u5bfc\u51fa\uff1a\u5c06\u6a21\u578b\u5bfc\u51fa\u4e3aONNX\u683c\u5f0f\uff0c\u518d\u4f7f\u7528ONNX Runtime\u52a0\u901f\u3002\u547d\u4ee4\u793a\u4f8b\uff1atorch.onnx.export(&#8230;)<\/li>\n<li>TensorRT\u4e0eFP16\uff1a\u5728GPU\u4e0a\u53ef\u5c06ONNX\u6a21\u578b\u7528TensorRT\u8f6c\u6362\u4e3aFP16\u6216INT8\u4ee5\u63d0\u9ad8\u901f\u5ea6\u5e76\u51cf\u5c11\u663e\u5b58\u3002<\/li>\n<li>\u6279\u5904\u7406\u4e0e\u5e8f\u5217\u5316\uff1a\u5bf9\u77ed\u6587\u672c\u5408\u5e76batch\u4ee5\u63d0\u9ad8GPU\u5229\u7528\u7387\uff0c\u4f46\u987b\u63a7\u5236\u6700\u5927\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n<h3>4. \u6784\u5efa\u63a8\u7406\u670d\u52a1<\/h3>\n<p>\u63a8\u8350\u4f7f\u7528\u8f7b\u91cf\u7684Web\u6846\u67b6\uff08\u5982Flask\u6216FastAPI\uff09\u914d\u5408Gunicorn\/Uvicorn\u7684\u8fdb\u7a0b\u7ba1\u7406\u5668\uff0c\u7ed3\u5408Nginx\u505a\u53cd\u5411\u4ee3\u7406\u4e0e\u8d1f\u8f7d\u5747\u8861\u3002\u5173\u952e\u70b9\uff1a<\/p>\n<ul>\n<li>\u8fdb\u7a0b\u4e0e\u7ebf\u7a0b\u914d\u7f6e\uff1a\u591a\u8fdb\u7a0b\u5bf9CPU\u7ed1\u5b9a\u7684\u6a21\u578b\u66f4\u53cb\u597d\uff1b\u5728GPU\u573a\u666f\u901a\u5e38\u4f7f\u7528\u5355\u8fdb\u7a0b\u591a\u7ebf\u7a0b\u6216gRPC\u4ee5\u907f\u514dCUDA\u4e0a\u4e0b\u6587\u7ade\u4e89\u3002<\/li>\n<li>\u8bf7\u6c42\u961f\u5217\u7ba1\u7406\uff1a\u8bbe\u7f6e\u5408\u7406\u7684\u8d85\u65f6\u4e0e\u961f\u5217\u957f\u5ea6\uff0c\u9632\u6b62\u7a81\u53d1\u6d41\u91cf\u5bfc\u81f4\u7cfb\u7edf\u5931\u7a33\u3002<\/li>\n<li>\u5065\u5eb7\u68c0\u67e5\u4e0e\u81ea\u52a8\u91cd\u542f\uff1a\u96c6\u6210systemd\u6216\u5bb9\u5668\u7f16\u6392\uff08\u5982Docker Compose \/ Kubernetes\uff09\u4fdd\u8bc1\u670d\u52a1\u53ef\u9760\u6027\u3002<\/li>\n<\/ul>\n<h2>\u6027\u80fd\u4f18\u5316\u4e0e\u76d1\u63a7\u7ec6\u8282<\/h2>\n<h3>CPU\u5c42\u9762\u7684\u4f18\u5316<\/h3>\n<ul>\n<li>\u542f\u7528\u591a\u7ebf\u7a0b\u6570\u5b66\u5e93\uff1a\u786e\u4fdd\u73af\u5883\u53d8\u91cf\u5982OMP_NUM_THREADS\u3001MKL_NUM_THREADS\u8bbe\u7f6e\u5f97\u5f53\uff08\u901a\u5e38\u4e0eCPU\u6838\u6570\u5339\u914d\u6216\u7a0d\u5c0f\uff09\u3002<\/li>\n<li>\u4f7f\u7528TorchScript\uff1a\u901a\u8fc7torch.jit.trace\/torch.jit.script\u5c06\u6a21\u578b\u7f16\u8bd1\u6210\u9759\u6001\u56fe\u4ee5\u51cf\u5c11Python\u5f00\u9500\u3002<\/li>\n<li>\u6a21\u578b\u88c1\u526a\/\u84b8\u998f\uff1a\u5c06BERT\u84b8\u998f\u4e3a\u66f4\u5c0f\u6a21\u578b\uff08DistilBERT\uff09\u5728\u4fdd\u6301\u7cbe\u5ea6\u7684\u540c\u65f6\u663e\u8457\u964d\u8d44\u6e90\u5360\u7528\u3002<\/li>\n<\/ul>\n<h3>GPU\u5c42\u9762\u7684\u4f18\u5316<\/h3>\n<ul>\n<li>\u4f7f\u7528TensorRT\u6216Apex\u7684\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\/\u63a8\u7406\uff0c\u964d\u4f4e\u5ef6\u8fdf\u4e0e\u663e\u5b58\u5360\u7528\u3002<\/li>\n<li>\u5408\u7406\u8bbe\u7f6ebatch size\u4ee5\u8fbe\u5230\u6700\u4f73\u541e\u5410\uff0c\u907f\u514d\u56e0\u8fc7\u5927batch\u5bfc\u81f4OOM\u3002<\/li>\n<li>\u76d1\u63a7CUDA\u663e\u5b58\u4e0eSM\u5229\u7528\u7387\uff0c\u8c03\u6574\u5e76\u53d1\u7ebf\u7a0b\u6216\u4f7f\u7528\u6a21\u578b\u5206\u7247\u3002<\/li>\n<\/ul>\n<h3>\u7f51\u7edc\u4e0e\u5ef6\u8fdf\u4f18\u5316<\/h3>\n<p>\u5bf9\u4e8e\u9762\u5411\u4e9a\u592a\u7528\u6237\u7684\u670d\u52a1\uff0c\u9999\u6e2fVPS\u5728\u7f51\u7edc\u5ef6\u8fdf\u4e0a\u901a\u5e38\u4f18\u4e8e\u5728\u5317\u7f8e\u7684\u7f8e\u56fdVPS\u6216\u7f8e\u56fd\u670d\u52a1\u5668\u3002\u5982\u679c\u7528\u6237\u57fa\u4e8e\u5168\u7403\u5206\u5e03\u5f0f\u8bbf\u95ee\uff0c\u53ef\u8003\u8651\u5728\u4e0d\u540c\u533a\u57df\u90e8\u7f72\u8f7b\u91cf\u63a8\u7406\u8282\u70b9\u5e76\u505a\u8def\u7531\u7b56\u7565\u3002\u4f7f\u7528CDN\u7f13\u5b58\u9759\u6001\u6216\u4f4e\u53d8\u5316\u7684\u63a8\u7406\u7ed3\u679c\u4e5f\u80fd\u6709\u6548\u964d\u4f4e\u540e\u7aef\u538b\u529b\u3002<\/p>\n<h3>\u76d1\u63a7\u4e0e\u81ea\u52a8\u6269\u5bb9<\/h3>\n<p>\u5efa\u8bae\u63a5\u5165Prometheus + Grafana\u76d1\u63a7\u6a21\u578b\u5ef6\u8fdf\u3001QPS\u3001CPU\/GPU\u4f7f\u7528\u7387\u4e0e\u5185\u5b58\u3002\u7ed3\u5408Horizontal Pod Autoscaler\uff08\u82e5\u4f7f\u7528Kubernetes\uff09\u6216\u81ea\u52a8\u6269\u5bb9\u811a\u672c\uff0c\u53ef\u5b9e\u73b0\u6309\u9700\u6269\u5bb9\uff0c\u4ee5\u5e94\u5bf9\u7a81\u53d1\u6d41\u91cf\u3002<\/p>\n<h2>\u5e94\u7528\u573a\u666f\u4e0e\u4f18\u52bf\u5bf9\u6bd4<\/h2>\n<h3>\u5178\u578b\u5e94\u7528\u573a\u666f<\/h3>\n<ul>\n<li>\u667a\u80fd\u5ba2\u670d\u4e0e\u95ee\u7b54\u7cfb\u7edf\uff1a\u8981\u6c42\u4f4e\u5ef6\u8fdf\u3001\u5e76\u53d1\u9ad8\uff0c\u9002\u5408GPU\u6216\u91cf\u5316\u7684CPU\u63a8\u7406\u3002<\/li>\n<li>\u5185\u5bb9\u5ba1\u6838\u4e0e\u6587\u672c\u5206\u7c7b\uff1a\u53ef\u6279\u91cf\u5904\u7406\uff0c\u9002\u5408\u5728\u591c\u95f4\u6216\u975e\u9ad8\u5cf0\u671f\u7528\u6279\u5904\u7406\u7b56\u7565\u3002<\/li>\n<li>\u641c\u7d22\u4e0e\u63a8\u8350\u4e2d\u7684\u8bed\u4e49\u7406\u89e3\uff1a\u9700\u8981\u8f83\u9ad8\u541e\u5410\u4e14\u6709\u8f83\u5f3a\u53ef\u6269\u5c55\u6027\u3002<\/li>\n<\/ul>\n<h3>\u9999\u6e2fVPS\u4e0e\u5176\u4ed6\u9009\u9879\u5bf9\u6bd4<\/h3>\n<p>\u9999\u6e2fVPS\u7684\u4f18\u52bf\u5728\u4e8e\u5bf9\u4e2d\u56fd\u3001\u4e1c\u5357\u4e9a\u7528\u6237\u63d0\u4f9b\u66f4\u4f4e\u7684\u7f51\u7edc\u5ef6\u8fdf\u4e0e\u66f4\u7a33\u5b9a\u7684\u8de8\u5883\u8bbf\u95ee\u4f53\u9a8c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u90e8\u7f72\u5728<span>\u7f8e\u56fdVPS<\/span>\u6216<span>\u7f8e\u56fd\u670d\u52a1\u5668<\/span>\u53ef\u80fd\u5728\u6210\u672c\u6216\u4e0e\u5317\u7f8e\u5ba2\u6237\u4e92\u52a8\u65f6\u6709\u4f18\u52bf\uff0c\u4f46\u5728\u5b9e\u65f6\u4ea4\u4e92\u573a\u666f\u5c06\u627f\u53d7\u66f4\u9ad8\u5ef6\u8fdf\u3002\u7efc\u5408\u8003\u8651\uff1a<\/p>\n<ul>\n<li>\u82e5\u76ee\u6807\u7528\u6237\u4e3b\u8981\u662f\u534e\u8bed\u6216\u4e9a\u592a\u5730\u533a\uff0c\u9999\u6e2fVPS\u901a\u5e38\u662f\u9996\u9009\u3002<\/li>\n<li>\u82e5\u76ee\u6807\u7528\u6237\u5206\u5e03\u5728\u5317\u7f8e\/\u6b27\u6d32\uff0c\u53ef\u8003\u8651\u5728\u5f53\u5730\u7684\u7f8e\u56fd\u670d\u52a1\u5668\u6216\u591a\u533a\u57df\u90e8\u7f72\u3002<\/li>\n<li>\u5bf9\u4e8e\u9884\u7b97\u654f\u611f\u4f46\u5ef6\u8fdf\u8981\u6c42\u4e0d\u9ad8\u7684\u4efb\u52a1\uff0c\u4f4e\u914d\u7f8e\u56fdVPS\u4e5f\u80fd\u6ee1\u8db3\u8bad\u7ec3\u6216\u79bb\u7ebf\u63a8\u7406\u9700\u6c42\u3002<\/li>\n<\/ul>\n<h2>\u9009\u8d2d\u5efa\u8bae\u4e0e\u90e8\u7f72\u6ce8\u610f\u4e8b\u9879<\/h2>\n<p>\u5728\u9009\u8d2d\u9999\u6e2fVPS\u65f6\uff0c\u8bf7\u5173\u6ce8\u4ee5\u4e0b\u8981\u70b9\uff1a<\/p>\n<ul>\n<li>CPU\u6838\u6570\u4e0e\u4e3b\u9891\uff1a\u9ad8\u4e3b\u9891\u5bf9\u5355\u8bf7\u6c42\u5ef6\u8fdf\u6709\u76f4\u63a5\u5f71\u54cd\uff1b\u591a\u6838\u5229\u4e8e\u5e76\u53d1\u541e\u5410\u3002<\/li>\n<li>\u5185\u5b58\u4e0e\u78c1\u76d8\uff1aBERT\u6a21\u578b\u52a0\u8f7d\u9700\u8981\u5145\u8db3\u5185\u5b58\uff1b\u63a8\u8350\u4f7f\u7528NVMe SSD\u4ee5\u52a0\u901f\u6a21\u578b\u52a0\u8f7d\u4e0e\u7f13\u5b58\u3002<\/li>\n<li>GPU\u652f\u6301\uff1a\u5982\u9700GPU\u63a8\u7406\uff0c\u786e\u8ba4VPS\u662f\u5426\u63d0\u4f9bNVIDIA\u663e\u5361\u4e0e\u76f8\u5e94\u9a71\u52a8\/\u865a\u62df\u5316\u652f\u6301\u3002<\/li>\n<li>\u7f51\u7edc\u4e0e\u5e26\u5bbd\uff1a\u82e5\u9700\u9891\u7e41\u62c9\u53d6\u6a21\u578b\u6216\u5904\u7406\u5927\u6279\u91cf\u8bf7\u6c42\uff0c\u9009\u62e9\u66f4\u9ad8\u5e26\u5bbd\u4e0e\u66f4\u4f4e\u65f6\u5ef6\u7684\u7f51\u7edc\u662f\u5fc5\u8981\u7684\u3002<\/li>\n<li>\u5b89\u5168\u4e0e\u5408\u89c4\uff1a\u5bf9\u4f01\u4e1a\u5e94\u7528\uff0c\u6ce8\u610f\u6570\u636e\u52a0\u5bc6\u3001\u8bbf\u95ee\u63a7\u5236\u4e0e\u5907\u4efd\u7b56\u7565\u3002<\/li>\n<\/ul>\n<p>\u90e8\u7f72\u6d41\u7a0b\u4e2d\u5e38\u89c1\u7684\u5751\uff1a<\/p>\n<ul>\n<li>\u5ffd\u89c6\u6570\u5b66\u5e93\u4e0e\u73af\u5883\u53d8\u91cf\u7684\u8bbe\u7f6e\uff0c\u5bfc\u81f4CPU\u5229\u7528\u7387\u4e0d\u9ad8\u6216\u7ebf\u7a0b\u4e89\u7528\u3002<\/li>\n<li>\u5728GPU\u4e0a\u76f2\u76ee\u589e\u52a0\u5e76\u53d1\u5bfc\u81f4CUDA OOM\u6216\u4e0a\u4e0b\u6587\u5207\u6362\u5e26\u6765\u6027\u80fd\u4e0b\u6ed1\u3002<\/li>\n<li>\u672a\u8bbe\u7f6e\u8bf7\u6c42\u9650\u6d41\u4e0e\u964d\u7ea7\u7b56\u7565\uff0c\u9762\u5bf9\u7a81\u53d1\u6d41\u91cf\u65f6\u53ef\u80fd\u5bfc\u81f4\u670d\u52a1\u4e0d\u53ef\u7528\u3002<\/li>\n<\/ul>\n<h2>\u603b\u7ed3<\/h2>\n<p>\u5728\u9999\u6e2fVPS\u4e0a\u90e8\u7f72BERT\uff0c\u4ece\u73af\u5883\u51c6\u5907\u3001\u6a21\u578b\u4f18\u5316\u5230\u670d\u52a1\u5316\u3001\u76d1\u63a7\u4e0e\u6269\u5bb9\uff0c\u90fd\u662f\u4e00\u5957\u7cfb\u7edf\u5de5\u7a0b\u3002<strong>\u5408\u7406\u9009\u62e9\u5b9e\u4f8b\u914d\u7f6e\uff08CPU\/GPU\/\u5185\u5b58\/\u7f51\u7edc\uff09\u5e76\u7ed3\u5408\u91cf\u5316\u3001ONNX\/TensorRT\u7b49\u4f18\u5316\u624b\u6bb5\uff0c\u80fd\u5728\u4fdd\u8bc1\u6a21\u578b\u6548\u679c\u7684\u524d\u63d0\u4e0b\u663e\u8457\u63d0\u5347\u6027\u80fd\u4e0e\u6210\u672c\u6548\u7387<\/strong>\u3002\u5982\u679c\u4f60\u7684\u76ee\u6807\u7528\u6237\u96c6\u4e2d\u5728\u4e9a\u592a\u5730\u533a\uff0c\u9999\u6e2f\u670d\u52a1\u5668\u80fd\u63d0\u4f9b\u66f4\u4f4e\u7684\u5ef6\u8fdf\u548c\u66f4\u597d\u7684\u7528\u6237\u4f53\u9a8c\uff1b\u82e5\u9762\u5411\u5168\u7403\u7528\u6237\uff0c\u53ef\u4ee5\u8003\u8651\u591a\u533a\u57df\uff08\u5305\u62ec\u7f8e\u56fdVPS\/\u7f8e\u56fd\u670d\u52a1\u5668\uff09\u8054\u5408\u90e8\u7f72\u4ee5\u5b9e\u73b0\u66f4\u5e7f\u8986\u76d6\u548c\u5bb9\u707e\u80fd\u529b\u3002<\/p>\n<p>\u5982\u679c\u4f60\u5e0c\u671b\u4ee5\u66f4\u5feb\u7684\u901f\u5ea6\u5728\u9999\u6e2fVPS\u4e0a\u5b8c\u6210BERT\u90e8\u7f72\u4e0e\u8fd0\u7ef4\uff0c\u6216\u60f3\u4e86\u89e3\u5177\u4f53\u7684\u9999\u6e2fVPS\u914d\u7f6e\u5efa\u8bae\u4e0e\u62a5\u4ef7\uff0c\u53ef\u53c2\u8003\u5e76\u8d2d\u4e70\u5408\u9002\u7684\u9999\u6e2fVPS\u4ea7\u54c1\uff1a<a href=\"https:\/\/www.server.hk\/cloud.php\">https:\/\/www.server.hk\/cloud.php<\/a>\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u60f3\u628aBERT\u6a21\u578b\u5728\u5927\u4e2d\u534e\u533a\u7a33\u5b9a\u4e0a\u7ebf\u5e76\u5b9e\u73b0\u4f4e\u5ef6\u8fdf\uff1f\u672c\u6587\u4ece\u539f\u7406\u3001\u90e8\u7f72\u6d41\u7a0b\u5230\u6027\u80fd\u4f18\u5316\u4e0e\u9009\u8d2d\u5efa\u8bae\uff0c\u624b\u628a\u624b\u6559\u4f60\u5728\u9999\u6e2fVPS \u90e8\u7f72 BERT\uff0c\u5feb\u901f\u83b7\u5f97\u9ad8\u541e\u5410\u4e0e\u4f4e\u5ef6\u8fdf\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4927],"tags":[2597,1595,543,720,747,2598,2599,527,107],"class_list":{"0":"post-209326","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-4927","7":"tag-bert","9":"tag-vps","10":"tag-720","11":"tag-747","12":"tag-2598","13":"tag-2599","14":"tag-527"},"_links":{"self":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts\/209326","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/comments?post=209326"}],"version-history":[{"count":1,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts\/209326\/revisions"}],"predecessor-version":[{"id":209327,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts\/209326\/revisions\/209327"}],"wp:attachment":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/media?parent=209326"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/categories?post=209326"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/tags?post=209326"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}