{"id":84635,"date":"2026-07-26T15:28:33","date_gmt":"2026-07-26T07:28:33","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/84635.html"},"modified":"2026-07-26T15:28:33","modified_gmt":"2026-07-26T07:28:33","slug":"onnx%e7%94%9f%e6%80%81%e7%ae%80%e4%bb%8b%e4%b8%8e%e5%ae%9e%e6%88%98%ef%bc%9aonnx-runtime","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/84635.html","title":{"rendered":"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime"},"content":{"rendered":"<h2>\u6982\u8ff0<\/h2>\n<p>\u5b98\u7f51&#xff0c;ONNX\u662fOpen Neural Network Exchange\u7684\u7f29\u5199&#xff0c;\u4f5c\u4e3a\u4e00\u79cd\u5f00\u653e\u5f00\u6e90&#xff08;GitHub&#xff0c;21.2K Star&#xff0c;4K Fork&#xff09;\u7684\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362\u683c\u5f0f&#xff0c;\u80fd\u591f\u5b9e\u73b0\u4e0d\u540c\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u95f4\u7684\u6a21\u578b\u8f6c\u6362\u4e0e\u5171\u4eab&#xff0c;\u4f7f\u5f97\u57fa\u4e8eONNX\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u591a\u79cd\u73af\u5883\u4e0b\u9ad8\u6548\u8fd0\u884c\u3002<\/p>\n<p>\u5728LLM\u4e4b\u524d&#xff08;\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3&#xff09;\u5c31\u662f\u975e\u5e38\u6d41\u884c\u7684\u6a21\u578b\u683c\u5f0f&#xff08;\u6a21\u578b\u901a\u7528\u8bed\u8a00&#xff09;&#xff0c;\u8de8\u6846\u67b6&#043;\u6a21\u578b\u7684\u4e2d\u95f4\u8868\u793a&#xff08;Intermediate Representation&#xff0c;IR&#xff09;\u683c\u5f0f\u3002<\/p>\n<p>\u6838\u5fc3\u7279\u6027&#xff1a;<\/p>\n<ul>\n<li>\u8de8\u6846\u67b6\u4e92\u64cd\u4f5c\u6027&#xff1a;\u5728PyTorch\u4e2d\u8bad\u7ec3\u7684\u6a21\u578b\u53ef\u4ee5\u5bfc\u51fa\u4e3aONNX\u683c\u5f0f&#xff0c;\u518d\u88abTensorFlow\u3001MXNet\u7b49\u5176\u4ed6\u6846\u67b6\u8bfb\u53d6<\/li>\n<li>\u6807\u51c6\u5316\u7b97\u5b50\u96c6&#xff1a;\u5b9a\u4e49\u7edf\u4e00\u7684\u6570\u5b66\u8fd0\u7b97\u548c\u795e\u7ecf\u7f51\u7edc\u64cd\u4f5c\u6807\u51c6<\/li>\n<li>Protocol Buffers\u5e8f\u5217\u5316&#xff1a;\u4f7f\u7528Google\u7684Protobuf\u4f5c\u4e3a\u5e95\u5c42\u5e8f\u5217\u5316\u534f\u8bae&#xff0c;.onnx\u6587\u4ef6\u5305\u542b\u5b8c\u6574\u7684\u8ba1\u7b97\u56fe\u7ed3\u6784\u548c\u6743\u91cd\u53c2\u6570<\/li>\n<li>\u7248\u672c\u5316\u7ba1\u7406&#xff1a;\u901a\u8fc7opset_version\u7ba1\u7406\u7b97\u5b50\u96c6\u7684\u6f14\u8fdb<\/li>\n<\/ul>\n<p>\u652f\u6301\u5404\u79cd\u4e0d\u540c\u683c\u5f0f\u7684\u6a21\u578b\u4e4b\u95f4\u7684\u4e92\u76f8\u8f6c\u6362&#xff1a;<\/p>\n<ul>\n<li>TensorFlow<\/li>\n<li>PyTorch<\/li>\n<li>TensorRT<\/li>\n<li>Keras<\/li>\n<li>CNTK<\/li>\n<li>Caffe<\/li>\n<li>TVM<\/li>\n<li>mxnet<\/li>\n<li>ncnn<\/li>\n<\/ul>\n<p>\u5b9a\u4f4d&#xff08;\u4e09\u5927\u76ee\u6807&#xff09;&#xff1a;<\/p>\n<ul>\n<li>\u5c06\u6a21\u578b\u4ece\u4efb\u4f55\u6846\u67b6\u8f6c\u6362\u4e3aONNX\u683c\u5f0f&#xff1b;<\/li>\n<li>\u5c06ONNX\u683c\u5f0f\u8f6c\u6362\u4e3a\u4efb\u4f55\u5176\u4ed6\u7684\u6846\u67b6&#xff1b;<\/li>\n<li>\u4f7f\u7528\u63a8\u7406\u5f15\u64ce\u8fd0\u884cONNX\u6a21\u578b&#xff0c;\u8ba9\u63a8\u7406\u66f4\u5feb\u901f\u3002<\/li>\n<\/ul>\n<p>\u751f\u6001\u56fe <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072831-6a65b71fed5c7.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b720140d3.jpg\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> \u5982\u4e0a\u56fe&#xff0c;\u5927\u6a21\u578b\u63a8\u7406\u7cfb\u7edf\u4e09\u4e2a\u7ec4\u6210\u90e8\u5206&#xff1a;<\/p>\n<ul>\n<li>\u786c\u4ef6&#xff1a;\u786c\u4ef6\u5e76\u884c\u5316&#xff0c;CPU\u3001GPU\u3001TPU\u5e95\u5c42\u52a0\u901f\u539f\u7406\u6709\u5f88\u5927\u5dee\u5f02&#xff0c;\u4f46\u90fd\u8ffd\u6c42\u6700\u5927\u5e76\u884c\u5316&#xff1b;<\/li>\n<li>\u8f6f\u4ef6&#xff1a;\u6a21\u578b\u672c\u8eab\u4e0d\u53d8\u5316&#xff0c;\u4f18\u5316\u8ba1\u7b97\u56fe\u7684\u4e24\u79cd\u65b9\u5411&#xff1a;\n<ul>\n<li>Low-level\u5e93&#xff1a;\u5982\u82f1\u4f1f\u8fbe\u63d0\u4f9b\u7684cuDNN\u5305\u53ef\u66f4\u597d\u5730\u5728GPU\u4e0a\u5904\u7406&#xff1b;<\/li>\n<li>\u56fe\u7f16\u8bd1\u5668&#xff1a;\u5982AI\u7f16\u8bd1\u5668TVM\u3001GLOW\u4e3b\u8981\u4f18\u5316\u524d\u5411\u4f20\u64ad\u6216\u540e\u5411\u4f20\u64ad\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u7b97\u6cd5&#xff1a;\u7b97\u6cd5\u7c7b\u65b9\u6cd5\u4f1a\u6539\u53d8\u6a21\u578b\u6216\u67b6\u6784\u6765\u52a0\u901f\u63a8\u7406&#xff0c;\u5982\u526a\u679d\u3001\u91cf\u5316\u3001\u84b8\u998f\u3001\u538b\u7f29\u3002<\/li>\n<\/ul>\n<p>\u5e94\u7528\u573a\u666f&#xff1a;<\/p>\n<ul>\n<li>\u79fb\u52a8\u7aef&#xff1a;\u5f88\u591a\u624b\u673aAPP\u540e\u53f0\u5c31\u662f\u901a\u8fc7ONNX\u628a\u5927\u6a21\u578b\u4ecePyTorch\u8f6c\u5230CoreML\u6216NNAPI&#xff1b;<\/li>\n<li>\u533b\u5b66\u5f71\u50cf&#xff1a;\u7528PyTorch\u8bad\u7ec3CT\/MRI\u7684\u75c5\u7076\u68c0\u6d4b\u6a21\u578b&#xff0c;ONNX&#043;OpenVINO\u53ef\u628a\u6a21\u578b\u90e8\u7f72\u5230\u666e\u901aCPU\u670d\u52a1\u5668&#xff1b;<\/li>\n<li>\u667a\u80fd\u5bb6\u5c45\u548cIoT&#xff1a;ONNX\u80fd\u628aPyTorch\u6a21\u578b\u8f6c\u6210\u9002\u5408ARM\u82af\u7247\u7684TensorRT\u5f15\u64ce&#xff1b;<\/li>\n<li>\u4e91\u7aef\u63a8\u7406\u670d\u52a1&#xff1a;HuggingFace\u3001Azure\u3001AWS\u90fd\u652f\u6301\u76f4\u63a5\u52a0\u8f7dONNX\u6a21\u578b&#xff0c;\u4e00\u952e\u90e8\u7f72\u4e3a\u4e91\u7aef\u63a8\u7406API\u3002<\/li>\n<\/ul>\n<p>\u4f18\u70b9<\/p>\n<ul>\n<li>\u8de8\u5e73\u53f0\u6027\u5f3a&#xff1a;PyTorch\u3001TensorFlow\u3001Scikit-learn\u3001XGBoost\u751a\u81f3\u90e8\u5206\u5927\u6a21\u578b\u90fd\u80fd\u5bfc\u51fa\u5230ONNX&#xff1b;<\/li>\n<li>\u90e8\u7f72\u7075\u6d3b&#xff1a;\u652f\u6301CPU\u3001GPU\u3001ARM\u3001FPGA\u3001NPU&#xff0c;\u57fa\u672c\u8986\u76d6\u4e3b\u6d41\u786c\u4ef6&#xff1b;<\/li>\n<li>\u63a8\u7406\u901f\u5ea6\u5feb&#xff1a;ONNX Runtime\u3001TensorRT\u7b49\u914d\u5408\u4e0b&#xff0c;\u80fd\u6bd4\u539f\u59cb\u6846\u67b6\u63d0\u53472~5\u500d\u901f\u5ea6&#xff1b;<\/li>\n<li>\u751f\u6001\u4e30\u5bcc&#xff1a;\u5fae\u8f6f\u3001AWS\u3001Intel\u3001NVIDIA\u7b49\u5de8\u5934\u90fd\u5728\u63a8&#xff0c;\u793e\u533a\u6d3b\u8dc3\u3002<\/li>\n<\/ul>\n<p>\u7f3a\u70b9<\/p>\n<ul>\n<li>\u4e0d\u652f\u6301\u6240\u6709\u7b97\u5b50&#xff1a;\u65b0\u6a21\u578b\u91cc\u7684\u81ea\u5b9a\u4e49\u5c42\u53ef\u80fd\u9700\u8981\u624b\u52a8\u5199\u63d2\u4ef6&#xff1b;<\/li>\n<li>\u8c03\u8bd5\u6210\u672c\u9ad8&#xff1a;\u6a21\u578b\u4e00\u65e6\u8f6c\u6362\u5931\u8d25&#xff0c;\u9519\u8bef\u4fe1\u606f\u53ef\u80fd\u6666\u6da9\u96be\u61c2&#xff1b;<\/li>\n<li>\u5927\u6a21\u578b\u9002\u914d\u6162&#xff1a;\u50cfGPT-4\u8fd9\u79cd\u8d85\u5927\u6a21\u578b&#xff0c;ONNX\u751f\u6001\u8fd8\u5728\u8ffd\u8d76&#xff1b;<\/li>\n<li>\u7248\u672c\u517c\u5bb9\u95ee\u9898&#xff1a;\u4e0d\u540c\u6846\u67b6\u6216opset\u53ef\u80fd\u5bfc\u81f4\u517c\u5bb9\u6027\u95ee\u9898\u3002<\/li>\n<\/ul>\n<p>\u8d8b\u52bf<\/p>\n<ul>\n<li>\u652f\u6301\u5927\u6a21\u578b&#xff1a;\u6b63\u5728\u9002\u914dTransformer\u3001LLM\u7684\u5e38\u89c1\u7b97\u5b50&#xff1b;<\/li>\n<li>\u8f7b\u91cf\u5316\u7ed3\u5408&#xff1a;\u652f\u6301\u91cf\u5316\u3001\u84b8\u998f&#xff0c;\u8ba9\u5927\u6a21\u578b\u80fd\u8dd1\u5728\u624b\u673a\u4e0a&#xff1b;<\/li>\n<li>\u63a8\u7406\u5f15\u64ce\u4e00\u4f53\u5316&#xff1a;\u672a\u6765\u53ef\u80fd\u76f4\u63a5\u4e00\u952e\u5bfc\u51fa&#043;\u81ea\u52a8\u4f18\u5316&#043;\u591a\u786c\u4ef6\u90e8\u7f72&#xff0c;\u6781\u5927\u964d\u4f4e\u95e8\u69db&#xff1b;<\/li>\n<li>\u4e91\u8fb9\u7aef\u534f\u540c&#xff1a;\u6210\u4e3aAI\u6a21\u578b\u5728\u4e91\u7aef\u4e0e\u8fb9\u7f18\u4e4b\u95f4\u7684\u6865\u6881&#xff0c;\u6253\u901a\u5168\u94fe\u8def\u3002<\/li>\n<\/ul>\n<h3>Converter<\/h3>\n<p>\u5373ONNX Converter&#xff0c;\u8f6c\u6362\u5668&#xff0c;\u7528\u4e8e\u5728\u4e0d\u540c\u6846\u67b6\u548cONNX\u4e4b\u95f4\u8f6c\u6362\u7684\u5de5\u5177&#xff1a;<\/p>\n<ul>\n<li>torch.onnx.export()&#xff1a;PyTorch\u8f6cONNX<\/li>\n<li>tf2onnx&#xff1a;TensorFlow\u8f6cONNX<\/li>\n<li>onnx2tf&#xff1a;ONNX\u8f6cTensorFlow<\/li>\n<li>onnx2pytorch&#xff1a;ONNX\u8f6cPyTorch<\/li>\n<li>onnx2keras&#xff1a;ONNX\u8f6cKeras<\/li>\n<\/ul>\n<h3>\u683c\u5f0f<\/h3>\n<p>ONNX\u4f7f\u7528Protocol Buffers\u683c\u5f0f\u5b58\u50a8\u6a21\u578b&#xff0c;\u76f4\u63a5\u5229\u7528Protobuf\u7684IR&#xff08;onnx.proto&#xff09;\u548c\u7f16\u8bd1\u5668\u80fd\u529b&#xff08;protoc&#xff09;&#xff0c;\u5c06\u6a21\u578b\u7684\u8ba1\u7b97\u56fe\u7ed3\u6784\u5e8f\u5217\u5316\u4e3a\u4e8c\u8fdb\u5236\u6587\u4ef6\u3002<\/p>\n<p>\u53ef\u76f4\u63a5\u4f7f\u7528Protobuf\u5c06ONNX\u6a21\u578b\u8f6c\u6362\u4e3a\u53ef\u8bfb\u7684\u6587\u672c&#xff1a;<\/p>\n<p><span class=\"token function\">wget<\/span> https:\/\/github.com\/onnx\/onnx\/raw\/main\/onnx\/onnx.proto<br \/>\n<span class=\"token function\">wget<\/span> https:\/\/github.com\/onnx\/models\/blob\/main\/validated\/vision\/classification\/mnist\/model\/mnist-8.onnx<br \/>\nprotoc <span class=\"token parameter variable\">&#8211;decode<\/span><span class=\"token operator\">&#061;<\/span>onnx.ModelProto <span class=\"token parameter variable\">&#8211;proto_path<\/span><span class=\"token operator\">&#061;<\/span>E:\/ipynb onnx.proto <span class=\"token operator\">&lt;<\/span> .\/mnist-8.onnx <span class=\"token operator\">&gt;<\/span> .\/mnist.onnx.txt<\/p>\n<p>\u5f97\u5230mnist.onnx.txt\u6587\u4ef6&#xff0c;\u4e3b\u8981\u7531\u4e09\u90e8\u5206\u7ec4\u6210&#xff1a;<\/p>\n<ul>\n<li>\u53ef\u6269\u5c55\u7684\u8ba1\u7b97\u56fe\u6a21\u578b\u7684\u5b9a\u4e49<\/li>\n<li>\u6807\u51c6\u6570\u636e\u7c7b\u578b\u5b9a\u4e49<\/li>\n<li>\u5185\u90e8\u7b97\u5b50\u5b9a\u4e49<\/li>\n<\/ul>\n<p>onnx.proto \u6587\u4ef6\u5b9a\u4e49Model\u3001Graph\u3001Node\u3001Tensor\u7b49\u6838\u5fc3\u6570\u636e\u7ed3\u6784&#xff0c;\u662f\u7406\u89e3\u548c\u64cd\u4f5cONNX\u6587\u4ef6\u7684\u5e95\u5c42\u57fa\u7840\u3002<\/p>\n<p>ModelProto<br \/>\n  \u251c\u2500\u2500 ir_version: \u7248\u672c\u53f7<br \/>\n  \u251c\u2500\u2500 opset_import: \u7b97\u5b50\u96c6\u7248\u672c<br \/>\n  \u251c\u2500\u2500 GraphProto<br \/>\n  \u2502\u251c\u2500\u2500 node[]: \u8ba1\u7b97\u8282\u70b9\u5217\u8868<br \/>\n  \u2502\u251c\u2500\u2500 initializer[]: \u6743\u91cd\u5f20\u91cf<br \/>\n  \u2502\u251c\u2500\u2500 input[]: \u8f93\u5165\u5b9a\u4e49<br \/>\n  \u2502\u2514\u2500\u2500 output[]: \u8f93\u51fa\u5b9a\u4e49<br \/>\n  \u2514\u2500\u2500 metadata_props: \u5143\u6570\u636e<\/p>\n<h3>\u91cf\u5316<\/h3>\n<p>ONNX\u652f\u6301\u591a\u79cd\u91cf\u5316\u65b9\u5f0f&#xff0c;\u5c06\u6a21\u578b\u4eceFP32\u538b\u7f29\u4e3aINT8\/FP16&#xff0c;\u51cf\u5c11\u6a21\u578b\u4f53\u79ef\u5e76\u63d0\u5347\u63a8\u7406\u901f\u5ea6&#xff1a;<\/p>\n<ul>\n<li>\u52a8\u6001\u91cf\u5316&#xff1a;\u63a8\u7406\u65f6\u52a8\u6001\u8ba1\u7b97\u91cf\u5316\u53c2\u6570<\/li>\n<li>\u9759\u6001\u91cf\u5316&#xff1a;\u4f7f\u7528\u6821\u51c6\u6570\u636e\u96c6\u9884\u8ba1\u7b97\u91cf\u5316\u53c2\u6570<\/li>\n<li>\u91cf\u5316\u611f\u77e5\u8bad\u7ec3&#xff08;QAT&#xff09;&#xff1a;\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6a21\u62df\u91cf\u5316\u6548\u679c<\/li>\n<\/ul>\n<p>\u53c2\u8003\u5927\u6a21\u578b\u57fa\u7840\u4e4b\u91cf\u5316\u3002<\/p>\n<h3>\u7b97\u5b50<\/h3>\n<p>\u53c2\u8003\u5b98\u65b9\u6587\u6863\u3002<\/p>\n<p>\u6bcf\u4e2aONNX\u7248\u672c\u5bf9\u5e94\u4e00\u4e2a\u7b97\u5b50\u96c6\u7248\u672c\u53f7opset_version\u3002\u968f\u7740\u7248\u672c\u6f14\u8fdb&#xff0c;\u65b0\u7b97\u5b50\u88ab\u52a0\u5165&#xff0c;\u8001\u7b97\u5b50\u88ab\u5e9f\u5f03\u3002\u5bfc\u51fa\u6a21\u578b\u65f6\u9700\u6307\u5b9aopset_version&#xff0c;\u51b3\u5b9a\u6a21\u578b\u4e2d\u53ef\u7528\u7684\u7b97\u5b50\u8303\u56f4\u3002<\/p>\n<ul>\n<li>Opset 11&#xff1a;\u5e38\u89c1\u57fa\u7840\u7b97\u5b50<\/li>\n<li>Opset 15&#xff1a;\u652f\u6301\u66f4\u591aTransformer\u76f8\u5173\u7b97\u5b50<\/li>\n<li>Opset 17-18&#xff1a;\u6269\u5c55\u66f4\u591a\u9ad8\u7ea7\u7b97\u5b50<\/li>\n<li>Opset 20&#043;&#xff1a;\u6700\u65b0\u7b97\u5b50\u96c6\u652f\u6301<\/li>\n<\/ul>\n<p>\u524d\u4e24\u8005\u7ed3\u5408&#xff0c;\u5373\u4e3aIR&#xff1b;\u7b97\u5b50\u5305\u542b\u57fa\u672c\u7b97\u5b50\u548c\u51fd\u6570\u7b97\u5b50Functions&#xff0c;\u540e\u8005\u8868\u793a\u590d\u5408\u64cd\u4f5c&#xff0c;\u7531\u5176\u4ed6\u7b97\u5b50\u7ec4\u5408\u800c\u6210\u7684\u4e00\u4e2a\u5b50\u56fe\u3002\u5982\u679c\u4e00\u4e2a\u8fd0\u884c\u65f6\u6ca1\u6709\u5b9e\u73b0\u67d0\u4e2a\u51fd\u6570\u7b97\u5b50&#xff0c;\u5c31\u5c06\u51fd\u6570\u7b97\u5b50\u66ff\u6362\u4e3a\u5176\u5b50\u56fe\u7684\u5177\u4f53\u5b9e\u73b0&#xff0c;\u79f0\u4e3a\u5185\u8054\u3002<\/p>\n<p>IR\u5b9a\u4e49\u8ba1\u7b97\u56fe\u7ed3\u6784&#xff1a; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b720297a4.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> \u8ba1\u7b97\u56fe\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u6709\u5411\u65e0\u73af\u56feDGA&#xff1a;<\/p>\n<ul>\n<li>\u8282\u70b9&#xff1a;\u8868\u793a\u5177\u4f53\u64cd\u4f5c&#xff0c;\u5982\u5377\u79ef\u3001Softmax\u7b97\u5b50&#xff1b;<\/li>\n<li>\u8fb9&#xff1a;\u8868\u793aTensor&#xff0c;\u5f20\u91cf&#xff1b;<\/li>\n<li>\u63a7\u5236\u6d41&#xff1a;\u7279\u6b8a\u8282\u70b9&#xff1b;<\/li>\n<li>\u8282\u70b9\u4f9d\u8d56\u5173\u7cfb&#xff1a;\u7279\u6b8a\u8fb9&#xff0c;\u8868\u793a\u8282\u70b9\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb<\/li>\n<\/ul>\n<p>\u7b97\u5b50\u64cd\u4f5c\u7b26\u7ec4\u6210&#xff1a;<\/p>\n<ul>\n<li>\u7b97\u5b50\u7c7b\u578b&#xff0c;\u5982Conv2D\u3001Softmax<\/li>\n<li>\u5c5e\u6027&#xff1a;\u7b97\u5b50\u53c2\u6570&#xff0c;\u5982\u5377\u79ef\u7684kernel_size\u3001stride\u3001paddin<\/li>\n<li>\u8f93\u5165\/\u8f93\u51fa\u5f20\u91cf\u4fe1\u606f&#xff1a;\u5982\u5f20\u91cf\u7684Shape\u3001DType<\/li>\n<\/ul>\n<p>model&#xff0c;graph&#xff0c;node&#xff0c;tensor\u7b49\u6570\u636e\u7ed3\u6784<\/p>\n<h3>\u8ba1\u7b97\u56fe<\/h3>\n<p>ONNX&#xff0c;\u5b9a\u4e49\u4e00\u5957\u6807\u51c6\u5316\u7684\u8ba1\u7b97\u56fe\u8868\u793a\u65b9\u5f0f&#xff0c;\u7528\u4e8e\u63cf\u8ff0\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u7684\u7ed3\u6784\u548c\u6743\u91cd\u3002<\/p>\n<p>PyTorch\u8bad\u7ec3\u7684\u6a21\u578b\u4f1a\u53d8\u6210\u4e00\u4e2a\u8ba1\u7b97\u56fe&#xff0c;\u7ee7\u800c\u5728\u5bf9\u5e94\u7684\u786c\u4ef6\u4e0a\u8fd0\u884c&#xff0c;\u5982\u679c\u8ba1\u7b97\u56fe\u53ef\u4ee5\u8868\u793a\u4e3a\u4e00\u4e2a\u4e2d\u95f4\u8868\u793aIR&#xff0c;\u5219\u56fe\u7f16\u8bd1\u5668&#xff08;\u6216AI\u7f16\u8bd1\u5668&#xff09;\u5c31\u5f88\u5bb9\u6613\u5728\u4e0d\u540c\u8bbe\u5907\u4e0a\u4f18\u5316&#xff0c;IR\u53ef\u4ee5\u7406\u89e3\u4e3a\u4e00\u79cd\u901a\u7528\u8868\u793a\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b720396ab.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> AI\u6846\u67b6\u548cAI\u7f16\u8bd1\u5668\u7684\u533a\u522b&#xff1a;<\/p>\n<ul>\n<li>AI\u6846\u67b6\u53ef\u80fd\u4f1a\u4f7f\u7528\u4e00\u4e9b\u65b0\u7684\u7b97\u5b50&#xff0c;\u4f46AI\u7f16\u8bd1\u5668\u53ef\u80fd\u6ca1\u6709\u5b9e\u73b0&#xff1b;<\/li>\n<li>\u4e00\u4e9b\u7b97\u5b50\u7684\u5b9e\u73b0\u65b9\u5f0f\u53ef\u80fd\u4e0d\u4e00\u6837&#xff1b;<\/li>\n<li>\u6709\u4e9b\u7f16\u8bd1\u5668\u53ea\u80fd\u548c\u67d0\u4e9b\u6846\u67b6\u9002\u914d&#xff0c;\u5982Intel\u7684OpenVino\u7f16\u8bd1\u5668\u53ea\u80fd\u7528\u4e8eTensorFlow&#xff0c;\u4e0d\u9002\u5408PyTorch\u3002<\/li>\n<\/ul>\n<p>ONNX\u662f\u4e00\u79cd\u901a\u7528\u6a21\u578b\u8ba1\u7b97\u56fe\u683c\u5f0f&#xff0c;\u5927\u90e8\u5206AI\u6846\u67b6\/AI\u7f16\u8bd1\u5668\u90fd\u652f\u6301\u5b83\u3002<\/p>\n<p>AI\u7f16\u8bd1\u5668\u5c06\u4e0d\u540cAI\u6846\u67b6\u4e2d\u7684\u9ad8\u7ea7\u8ba1\u7b97\u56fe\u6620\u5c04\u4e3a\u7279\u5b9a\u786c\u4ef6\u4e0a\u7684\u6267\u884c\u64cd\u4f5c&#xff0c;\u4f1a\u8fdb\u884c\u4e00\u7cfb\u5217\u4f18\u5316&#xff1a;<\/p>\n<ul>\n<li>\u56fe\u91cd\u5199&#xff1a;\u56fe\u7ed3\u6784\u51b3\u5b9a\u64cd\u4f5cOP\u7684\u6267\u884c\u987a\u5e8f&#xff0c;\u800c\u4f5c\u4e1a\u8c03\u5ea6&#xff08;Job scheduling&#xff09;\u8003\u8651\u5982\u4f55\u5c06OP\u6267\u884c\u987a\u5e8f\u6700\u4f73\u5316&#xff1a;\n<ul>\n<li>\u5220\u9664\u4e00\u4e9b\u8282\u70b9\u6216\u8fb9<\/li>\n<li>\u7b97\u5b50\u878d\u5408<\/li>\n<li>\u5b50\u56fe\u66ff\u4ee3<\/li>\n<li>\u5220\u9664\u65e0\u7528\u7684\u5c42<\/li>\n<\/ul>\n<\/li>\n<li>\u7b97\u5b50\u878d\u5408&#xff1a;\u5c06\u591a\u4e2a\u8fde\u7eed\u7684\u64cd\u4f5c\u5408\u5e76\u4e3a\u4e00\u4e2a\u590d\u5408\u64cd\u4f5c&#xff0c;\u5728\u8ba1\u7b97\u56fe\u4e2d&#xff0c;\u7531\u5f88\u591a\u7ec6\u7c92\u5ea6\u7684\u7b97\u5b50\u7ec4\u6210&#xff0c;\u6bcf\u4e2a\u64cd\u4f5c\u90fd\u8981\u72ec\u7acb\u7684\u8ba1\u7b97\u548c\u8bfb\u5199\u5185\u5b58&#xff0c;\u53ef\u4ee5\u5c06\u591a\u4e2a\u8fde\u7eed\u7684\u64cd\u4f5c\u5408\u5e76\u4e3a\u4e00\u4e2a\u590d\u5408\u64cd\u4f5c&#xff0c;\u8fd9\u6837\u53ef\u4ee5\u51cf\u5c11\u4e2d\u95f4\u7ed3\u679c\u7684\u5b58\u50a8\u548c\u8bfb\u53d6&#xff0c;\u4e5f\u80fd\u5145\u5206\u5229\u7528\u786c\u4ef6\u4e13\u6709\u7684Kernel&#xff08;\u5982\u4e13\u4e3a\u77e9\u9635\u4e58\u64cd\u4f5c\u7684Tensor Core&#xff09;\u3002\u6bd4\u5982&#xff0c;convolution\u3001ReLU\u3001BatchNorm\u53ef\u4ee5\u878d\u5408\u4e3a\u8ba1\u7b97\u56fe\u4e2d\u7684\u4e00\u4e2a\u8282\u70b9&#xff08;\u7b97\u5b50&#xff09; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b7207ed6d.jpg\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<li>\u7b97\u5b50\u5206\u914d\u4e0e\u8c03\u5ea6\u4f18\u5316 \u4e0a\u8ff0\u4e24\u4e2a\u4f18\u5316\u57fa\u672c\u4e0a\u662f\u548c\u786c\u4ef6\u65e0\u5173\u7684&#xff0c;\u5bf9\u4e8e\u56fe\u7f16\u8bd1\u5668\u6765\u8bf4&#xff0c;\u6700\u7ec8\u64cd\u4f5c\u8981\u8fd0\u884c\u5728\u5404\u79cd\u786c\u4ef6\u4e0a&#xff0c;\u5b83\u5145\u5f53\u4e86\u786c\u4ef6\u7684\u62bd\u8c61&#xff0c;\u9009\u62e9\u4e0d\u540c\u7684\u7b97\u5b50\u8fd0\u884c\u5728\u7279\u5b9a\u786c\u4ef6\u4e0a&#xff08;GPU\u3001NPU\u3001CPU&#xff09;&#xff0c;\u6bd4\u5982\u5377\u79ef\u64cd\u4f5c\u9002\u5408 GPU\u8fd0\u884c&#xff0c;\u6807\u91cf\u8fd0\u7b97\u9002\u5408CPU\u8fd0\u884c&#xff1b;\u901a\u8fc7\u8003\u8651\u8de8\u786c\u4ef6\u4f9d\u8d56\u5173\u7cfb&#xff0c;\u5408\u7406\u8c03\u5ea6\u6267\u884c\u7684\u987a\u5e8f&#xff0c;\u6700\u7ec8\u63d0\u5347\u8fd0\u884c\u6548\u7387\u3002<\/li>\n<\/ul>\n<p>\u76ee\u524d\u5927\u90e8\u5206 AI \u6846\u67b6\u90fd\u6709\u81ea\u5df1\u7684\u8ba1\u7b97\u56fe\u8868\u793a\u65b9\u6cd5\u548c\u4f18\u5316\u76ee\u6807&#xff0c;\u53ef\u4ee5\u6839\u636e\u9700\u6c42\u5f00\u53d1\u5927\u6a21\u578b&#xff0c;\u4f46\u5927\u6a21\u578b\u90e8\u7f72\u8d8a\u6765\u8d8a\u591a\u6837\u5316&#xff08;\u6bd4\u5982\u8fb9\u7f18\u8bbe\u5907\u8fd0\u884c&#xff0c;\u4e00\u4e2a\u6846\u67b6\u90e8\u7f72\u5230\u53e6\u5916\u4e00\u4e2a\u6846\u67b6\u652f\u6301\u7684\u73af\u5883&#xff09;&#xff0c;\u90a3\u4e48\u4e0d\u540c\u6846\u67b6\u5f00\u53d1\u51fa\u6765\u7684\u6a21\u578b\u8de8\u5e73\u53f0\u90e8\u7f72\u6216\u6027\u80fd\u4f18\u5316\u5c31\u975e\u5e38\u56f0\u96be\u4e86\u3002<\/p>\n<p>\u8fd9\u65f6\u5019\u8ba1\u7b97\u56fe\u5982\u679c\u6709\u4e00\u4e2a\u7edf\u4e00\u7684\u4e2d\u95f4\u8868\u793a IR&#xff0c;\u7528\u4e8e\u5c4f\u853d\u4e0d\u540c\u6846\u67b6\u4e4b\u95f4\u7684\u5dee\u5f02&#xff0c;\u4ece\u800c\u8ba9\u4e0d\u540c AI \u7f16\u8bd1\u5668\u53bb\u4f18\u5316&#xff0c;ONNX \u5c31\u662f\u8fd9\u6837\u4e00\u4e2a\u901a\u7528 IR\u3002<\/p>\n<h2>\u5b9e\u6218<\/h2>\n<h3>\u8f6c\u6362<\/h3>\n<p>PyTorch\u8f6c\u6362\u4e3aONNX&#xff1a;<\/p>\n<p><span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">import<\/span> torch<span class=\"token punctuation\">.<\/span>nn <span class=\"token keyword\">as<\/span> nn<\/p>\n<p><span class=\"token comment\"># \u5b9a\u4e49\u6a21\u578b<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">SimpleClassifier<\/span><span class=\"token punctuation\">(<\/span>nn<span class=\"token punctuation\">.<\/span>Module<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n<span class=\"token keyword\">def<\/span> <span class=\"token function\">__init__<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n<span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>SimpleClassifier<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nself<span class=\"token punctuation\">.<\/span>fc <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span><span class=\"token number\">768<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># BERT\u8f93\u51fa768\u7ef4<\/span><br \/>\n<span class=\"token keyword\">def<\/span> <span class=\"token function\">forward<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> x<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n<span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>fc<span class=\"token punctuation\">(<\/span>x<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># \u521d\u59cb\u5316\u6a21\u578b<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> SimpleClassifier<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nmodel<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\ndummy_input <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>randn<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">768<\/span><span class=\"token punctuation\">)<\/span><br \/>\ntorch<span class=\"token punctuation\">.<\/span>onnx<span class=\"token punctuation\">.<\/span>export<span class=\"token punctuation\">(<\/span><br \/>\nmodel<span class=\"token punctuation\">,<\/span><br \/>\ndummy_input<span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token string\">&#034;classifier.onnx&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\ninput_names<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;input&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\noutput_names<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;output&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\ndynamic_axes<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;input&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;batch_size&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;output&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;batch_size&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\nopset_version<span class=\"token operator\">&#061;<\/span><span class=\"token number\">17<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u89e3\u8bfb&#xff1a;<\/p>\n<ul>\n<li>dynamic_axes&#xff1a;\u4fdd\u8bc1\u8f93\u5165\u8f93\u51fa\u652f\u6301\u52a8\u6001Batch&#xff1b;<\/li>\n<li>opset_version&#xff1a;\u7248\u672c\u53f7&#xff0c;\u5efa\u8bae\u900913&#043;&#xff0c;\u517c\u5bb9\u6027\u66f4\u597d\u3002<\/li>\n<\/ul>\n<h3>\u63a8\u7406<\/h3>\n<p>\u4f7f\u7528ONNX Runtime&#xff1a;<\/p>\n<p><span class=\"token keyword\">import<\/span> onnxruntime <span class=\"token keyword\">as<\/span> ort<br \/>\n<span class=\"token keyword\">import<\/span> numpy <span class=\"token keyword\">as<\/span> np<\/p>\n<p><span class=\"token comment\"># \u52a0\u8f7d\u6a21\u578b<\/span><br \/>\nsession <span class=\"token operator\">&#061;<\/span> ort<span class=\"token punctuation\">.<\/span>InferenceSession<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;classifier.onnx&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\ninput_data <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>randn<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">768<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>astype<span class=\"token punctuation\">(<\/span>np<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">)<\/span><br \/>\ninputs <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;input&#034;<\/span><span class=\"token punctuation\">:<\/span> input_data<span class=\"token punctuation\">}<\/span><br \/>\n<span class=\"token comment\"># \u63a8\u7406<\/span><br \/>\noutputs <span class=\"token operator\">&#061;<\/span> session<span class=\"token punctuation\">.<\/span>run<span class=\"token punctuation\">(<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">,<\/span> inputs<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u63a8\u7406\u7ed3\u679c:&#034;<\/span><span class=\"token punctuation\">,<\/span> outputs<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h3>FastAPI<\/h3>\n<p>\u96c6\u6210FastAPI&#xff0c;\u90e8\u7f72\u4e3a\u670d\u52a1&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> fastapi <span class=\"token keyword\">import<\/span> FastAPI<br \/>\n<span class=\"token keyword\">import<\/span> uvicorn<\/p>\n<p>app <span class=\"token operator\">&#061;<\/span> FastAPI<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nsession <span class=\"token operator\">&#061;<\/span> ort<span class=\"token punctuation\">.<\/span>InferenceSession<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;classifier.onnx&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token decorator annotation punctuation\">&#064;app<span class=\"token punctuation\">.<\/span>post<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\/predict&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">def<\/span> <span class=\"token function\">predict<\/span><span class=\"token punctuation\">(<\/span>data<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">list<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\ninput_data <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>array<span class=\"token punctuation\">(<\/span>data<span class=\"token punctuation\">,<\/span> dtype<span class=\"token operator\">&#061;<\/span>np<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>reshape<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><br \/>\nresult <span class=\"token operator\">&#061;<\/span> session<span class=\"token punctuation\">.<\/span>run<span class=\"token punctuation\">(<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;input&#034;<\/span><span class=\"token punctuation\">:<\/span> input_data<span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">return<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;prediction&#034;<\/span><span class=\"token punctuation\">:<\/span> result<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>tolist<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">}<\/span><br \/>\n<span class=\"token keyword\">if<\/span> __name__ <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token string\">&#034;__main__&#034;<\/span><span class=\"token punctuation\">:<\/span><br \/>\nuvicorn<span class=\"token punctuation\">.<\/span>run<span class=\"token punctuation\">(<\/span>app<span class=\"token punctuation\">,<\/span> host<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;0.0.0.0&#034;<\/span><span class=\"token punctuation\">,<\/span> port<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8000<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h2>ONNX Runtime<\/h2>\n<p>\u5b98\u7f51&#xff0c;\u7b80\u79f0ORT&#xff0c;\u5fae\u8f6f\u5b98\u65b9\u5f00\u6e90&#xff08;GitHub&#xff0c;21.2K Star&#xff0c;4.1K Fork&#xff09;\u8de8\u5e73\u53f0\u3001\u9ad8\u6027\u80fd\u8fd0\u884c\u65f6\u63a8\u7406\u5f15\u64ce&#xff0c;\u652f\u6301CPU\/GPU\/NPU&#xff0c;\u4e13\u95e8\u7528\u4e8e\u6267\u884cONNX\u6a21\u578b&#xff0c;\u8d1f\u8d23\u5c06ONNX\u683c\u5f0f\u7684\u6a21\u578b\u52a0\u8f7d\u5230\u5185\u5b58\u4e2d&#xff0c;\u5e76\u5229\u7528\u786c\u4ef6\u52a0\u901f\u80fd\u529b\u9ad8\u6548\u8fd0\u884c\u63a8\u7406\u3002<\/p>\n<p>\u6838\u5fc3\u7279\u6027&#xff1a;<\/p>\n<ul>\n<li>\u591a\u8bed\u8a00API&#xff1a;\u652f\u6301Python\u3001C&#043;&#043;\u3001C#\u3001Java\u3001JavaScript\u3001Rust<\/li>\n<li>\u8de8\u5e73\u53f0\u652f\u6301&#xff1a;Linux\u3001Windows\u3001macOS\u3001Android\u3001iOS&#xff0c;\u751a\u81f3Web\u6d4f\u89c8\u5668<\/li>\n<li>Execution Providers\u673a\u5236&#xff1a;\u901a\u8fc7\u63d2\u4ef6\u5f0f\u7684\u6267\u884c\u63d0\u4f9b\u8005&#xff0c;\u652f\u6301\u4e0d\u540c\u786c\u4ef6\u540e\u7aef<\/li>\n<li>\u56fe\u4f18\u5316\u6280\u672f&#xff1a;\u81ea\u52a8\u8fdb\u884c\u7b97\u5b50\u878d\u5408\u3001\u5e38\u91cf\u6298\u53e0\u3001\u5185\u5b58\u4f18\u5316\u7b49<\/li>\n<li>\u8bad\u7ec3\u52a0\u901f&#xff1a;\u9664\u4e86\u63a8\u7406&#xff0c;\u8fd8\u652f\u6301\u5927\u6a21\u578b\u8bad\u7ec3\u52a0\u901f<\/li>\n<\/ul>\n<p>\u6838\u5fc3\u6982\u5ff5<\/p>\n<ul>\n<li>\u8fd0\u884c\u65f6&#xff1a;ORT\u4e4b\u6240\u4ee5\u53ef\u5728\u4e0d\u540c\u7684\u786c\u4ef6\u8bbe\u5907\u4e0a\u8fd0\u884c&#xff0c;\u62bd\u8c61\u4e00\u4e2aAPI&#xff0c;\u79f0\u4e3aExecution Providers&#xff08;EP&#xff09;&#xff0c;\u5982CUDA EP\u3002<\/li>\n<li>EP&#xff1a;EP \u662f\u5bf9\u67d0\u4e2a\u786c\u4ef6\u7684\u4e00\u79cd\u62bd\u8c61&#xff0c;\u6bcf\u4e2a EP \u5411 ORT \u8bf4\u660e\u5176\u786c\u4ef6\u652f\u6301\u7684\u80fd\u529b&#xff0c;\u6bd4\u5982\u652f\u6301\u7684\u7b97\u5b50\u3001GPU\u663e\u5b58\u7ba1\u7406\u3001\u5f02\u6784\u8c03\u5ea6\u6267\u884c\u3002EP\u4e00\u822c\u7531ORT\u5b98\u65b9\u548c\u5382\u5546\u5408\u4f5c\u5f00\u53d1&#xff0c;\u5982cuDNN\u8c03\u7528\u3002\u5982\u679c\u6ca1\u6709\u7b97\u5b50\u5206\u914d\u7ed9 EP \u6267\u884c&#xff0c;\u90a3 ORT \u4f1a\u6709\u56de\u9000\u673a\u5236&#xff0c;\u5982\u8ba9CPU\u53bb\u6267\u884c&#xff0c;\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u603b\u662f\u80fd\u591f\u8fd0\u884c\u3002\u4e24\u79cd\u7c7b\u578b&#xff1a;\n<ul>\n<li>Kernel EP&#xff0c;\u5982CPUExecutionProvider\u3001CUDAExecutionProvider&#xff0c;ORT\u5b98\u65b9\u5b9e\u73b0\u7684&#xff1b;<\/li>\n<li>Runtime EP&#xff0c;\u4e00\u822c\u662f\u786c\u4ef6\u5382\u5546\u5b9e\u73b0\u7684&#xff0c;\u5982TensorRTExecutionProvider&#xff0c;\u7528\u4e8e\u4e13\u95e8\u7684\u4f18\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n<li>Graph\u56fe\u4f18\u5316&#xff1a;ONNX\u6a21\u578b\u5c31\u662f\u4e00\u4e2a\u8ba1\u7b97\u56fe&#xff0c;ORT\u5c31\u662f\u5bf9\u8ba1\u7b97\u56fe\u8fdb\u884c\u4f18\u5316\u3002\u56fe\u4f18\u5316\u5206\u4e3a\u79bb\u7ebf\u548c\u5728\u7ebf\u6a21\u5f0f&#xff1b;\u4e5f\u53ef\u5206\u4e3a\u4e09\u5c42\u4f18\u5316&#xff1a;\n<ul>\n<li>\u57fa\u7840\u4f18\u5316&#xff1a;\u4fdd\u7559\u8bed\u4e49\u7684\u56fe\u91cd\u5199&#xff0c;\u548c\u5177\u4f53\u540e\u7aef\u786c\u4ef6\u65e0\u5173&#xff0c;\u5728\u56fe\u5206\u533a\u4e4b\u524d\u8fd0\u884c&#xff0c;\u5305\u62ec\u5e38\u91cf\u6298\u53e0\u3001\u5197\u4f59\u8282\u70b9\u5220\u9664&#xff08;\u5982Identity\u3001Slice\u3001Unsqueeze\u3001Dropout&#xff09;\u3001\u7b97\u5b50\u878d\u5408\u3002<\/li>\n<li>\u6269\u5c55\u4f18\u5316&#xff1a;\u5305\u62ec\u66f4\u590d\u6742\u7684\u8282\u70b9\u96c6\u6210&#xff0c;\u53ea\u9002\u7528\u4e8e\u5206\u914d\u7ed9CPU\u3001CUDA\u548cROCm\u6267\u884c\u7684\u8282\u70b9&#xff0c;\u4e5f\u5c31\u662f\u8bf4\u548c\u786c\u4ef6\u6709\u5173&#xff0c;\u6216\u8005\u548c\u7279\u5b9a\u6a21\u578b\u67b6\u6784\u6709\u5173\u3002<\/li>\n<li>\u5f20\u91cf\u5185\u5b58\u5e03\u5c40\u4f18\u5316&#xff1a;\u6539\u53d8\u6570\u636e\u5e03\u5c40&#xff0c;\u8ba9\u786c\u4ef6\u66f4\u597d\u4f18\u5316<\/li>\n<\/ul>\n<\/li>\n<li>\u56fe\u8f6c\u6362&#xff1a;\u56fe\u4f18\u5316\u53ef\u4ece\u5176\u4ed6\u89d2\u5ea6\u7406\u89e3&#xff0c;\u5982\u4ece\u6a21\u578b\u7684\u5e94\u7528\u6c34\u5e73\u5206\u6790\u770b&#xff0c;\u4f18\u5316\u4e00\u822c\u88ab\u79f0\u4e3a\u56fe\u8f6c\u6362&#xff0c;\u56fe\u8f6c\u6362\u5305\u542b\u56fe\u4f18\u5316\u3002\u5206\u4e3a\u4e09\u4e2a\u5c42\u6b21&#xff1a;\n<ul>\n<li>\u65e0\u6761\u4ef6\u8f6c\u6362&#xff1a;\u8ba1\u7b97\u56fe\u751f\u6210\u540e\u53ef\u4ee5\u7acb\u523b\u8fdb\u884c\u7684\u56fe\u8f6c\u6362&#xff0c;\u5982Cast\u3001Memcpy\u7b49\u64cd\u4f5c\u786e\u4fdd\u5f20\u91cf\u7c7b\u578b\u548c\u8bbe\u5907\u4e00\u81f4<\/li>\n<li>\u901a\u7528\u8f6c\u6362&#xff1a;\u548c\u8bbe\u5907\u65e0\u5173\u7684\u8f6c\u6362&#xff0c;\u6bd4\u5982\u5728\u63a8\u7406\u9636\u6bb5\u79fb\u9664\u8ba1\u7b97\u56fe\u4e2d\u7684Dropout&#xff1b;<\/li>\n<li>EP\u7ea7\u522b\u8f6c\u6362&#xff1a;\u56fe\u8f6c\u6362\u4e00\u79cd\u5206\u7c7b&#xff1a;\n<ul>\n<li>\u5168\u5c40\u6a21\u5f0f&#xff1a;\u56fe\u8f6c\u6362\u4f1a\u5bf9\u6574\u4e2a\u8ba1\u7b97\u56fe\u8fdb\u884c\u8f6c\u6362&#xff0c;\u5728ORT\u4e2d\u8be5\u63a5\u53e3\u79f0\u4e3aGraph Transformer&#xff1b;<\/li>\n<li>\u5c40\u90e8\u6a21\u5f0f&#xff1a;\u56fe\u8f6c\u6362\u5bf9\u67d0\u4e2a\u5b50\u56fe\u6216\u4e00\u4e9b\u8282\u70b9\u8fdb\u884c\u8f6c\u6362&#xff0c;\u5728ORT\u4e2d\u8be5\u63a5\u53e3\u79f0\u4e3aRewriting Rule\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<li>Graph Partitioning&#xff1a;\u56fe\u5206\u533a\u3002\u548cEP\u6709\u5173&#xff0c;ORT\u6839\u636e\u53ef\u7528EP\u5c06\u6a21\u578b\u56fe\u5212\u5206\u4e3a\u591a\u4e2a\u5b50\u56fe&#xff08;\u6bcf\u4e2a\u5b50\u56fe\u662f\u67d0\u4e2aEP\u8fde\u7eed\u80fd\u5904\u7406\u7684\u8282\u70b9&#xff09;&#xff0c;\u6bcf\u4e2a\u5b50\u56fe\u5bf9\u5e94\u4e00\u4e2a\u4e0d\u540c\u7684EP&#xff0c;ORT\u63d0\u4f9b\u4e00\u4e2a\u9ed8\u8ba4EP&#xff08;cpu&#xff09;&#xff0c;\u5982\u679c\u6ca1\u6709\u627e\u5230\u66f4\u9ad8\u6548\u7684EP\u5904\u7406\u5b50\u56fe&#xff0c;\u5c31\u4f7f\u7528\u9ed8\u8ba4\u7684EP\u5904\u7406\u3002 \u56fe\u5206\u533a\u7684\u5177\u4f53\u6280\u672f\u5b9e\u9645\u4e0a\u4e5f\u5f88\u7b80\u5355&#xff0c;\u6309\u7167\u7279\u5b9a\u987a\u5e8f\u8003\u8651\u53ef\u7528\u7684EP&#xff0c;\u4e3a\u6bcf\u4e2aEP\u5206\u914d\u5176\u80fd\u591f\u5904\u7406\u7684\u6700\u5927\u5b50\u56fe&#xff0c;\u91c7\u7528\u8d2a\u5a6a\u641c\u7d22\u7684\u6a21\u5f0f\u3002 \u6700\u7ec8\u56fe\u8f6c\u6362\u548c\u56fe\u5206\u533a\u5c06\u539f\u59cb\u6a21\u578b\u56fe\u8f6c\u6362\u4e3a\u7531\u5206\u914d\u7ed9\u9ed8\u8ba4EP\u6216\u5176\u4ed6\u5df2\u6ce8\u518cEP\u7684\u7b97\u5b50\u7ec4\u6210\u7684\u65b0\u56fe&#xff0c;ORT\u6267\u884c\u5f15\u64ce\u8d1f\u8d23\u8fd0\u884c\u8be5\u56fe&#xff0c;\u4e5f\u53ef\u4ee5\u53eb\u505a\u56fe\u8c03\u5ea6\u548c\u6267\u884c\u3002<\/li>\n<li>\u6574\u4f53\u67b6\u6784 \u4ece\u7ec4\u4ef6\u7684\u89d2\u5ea6\u7406\u89e3&#xff0c;\u5982\u4e0b\u56fe&#xff1a; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b72097764.jpg\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> \u6574\u4f53\u6bd4\u8f83\u7b80\u5355&#xff1a;\n<ul>\n<li>\u52a0\u8f7dONNX\u6a21\u578b&#xff1b;<\/li>\n<li>\u8f6c\u6362\u6210\u5185\u5b58\u4e2d\u7684\u56fe\u7ed3\u6784&#xff1b;<\/li>\n<li>\u56fe\u5206\u533a&#xff0c;\u6839\u636e\u652f\u6301\u60c5\u51b5\u62c6\u5206\u5b50\u56fe&#xff1b;<\/li>\n<li>\u5e76\u884c\u8c03\u5ea6\u5668\u8fd0\u884c\u5404\u4e2a\u5b50\u56fe&#xff0c;\u4ea4\u7531\u4e0d\u540cEP\u6267\u884c \u4ece\u56fe\u4f18\u5316\u548c\u56fe\u5206\u533a\u5904\u7406\u6d41\u7a0b\u89d2\u5ea6\u7406\u89e3&#xff0c;\u66f4\u8be6\u7ec6&#xff1a; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072832-6a65b720b9e2a.jpg\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/> \u6db5\u76d6\u6838\u5fc3\u6d41\u7a0b&#xff1a;<\/li>\n<li>ONNX\u6a21\u578b\u52a0\u8f7d\u6210\u4e2d\u95f4\u8868\u793aIR&#xff1b;<\/li>\n<li>\u8fdb\u884c\u5e73\u53f0\u65e0\u5173\u7684\u56fe\u4f18\u5316&#xff08;\u5982\u8282\u70b9\u878d\u5408&#xff09;<\/li>\n<li>\u8fdb\u884c\u56fe\u5206\u533a&#xff1b;<\/li>\n<li>\u9488\u5bf9\u6bcf\u4e2aEP\u505a\u5b9a\u5236\u4f18\u5316&#xff1b;<\/li>\n<li>\u4f7f\u7528\u987a\u5e8f\u6216\u5e76\u884c\u8c03\u5ea6\u6267\u884cEP&#xff1b;<\/li>\n<li>\u5404\u4e2aEP\u8d1f\u8d23\u5177\u4f53\u6267\u884c\u5b50\u56fe<\/li>\n<\/ul>\n<\/li>\n<li>ORT\u4e3b\u8981\u76ee\u6807&#xff1a;\n<ul>\n<li>\u5145\u5206\u4f7f\u7528\u786c\u4ef6\u80fd\u529b&#xff1a;\u6700\u5927\u9650\u5ea6\u5730\u81ea\u52a8\u5229\u7528\u4e0d\u540c\u786c\u4ef6\u5e73\u53f0\u4e0a\u7684\u52a0\u901f\u5668&#xff08;\u5982GPU&#xff09;\u548c\u8fd0\u884c\u65f6&#xff1b;<\/li>\n<li>\u63d0\u4f9b\u786c\u4ef6\u5de5\u4f5c\u7684\u62bd\u8c61\u63a5\u53e3&#xff1a;\u4e3a\u81ea\u5b9a\u4e49\u52a0\u901f\u5668\u548c\u8fd0\u884c\u65f6\u63d0\u4f9b\u6b63\u786e\u7684\u62bd\u8c61\u548c\u8fd0\u884c\u65f6\u652f\u6301&#xff0c;\u8fd9\u79cd\u62bd\u8c61\u4e3aEP&#xff0c;\u5b83\u5b9a\u4e49\u5e76\u5411ORT\u516c\u5f00\u529f\u80fd&#xff1b;<\/li>\n<li>\u517c\u5bb9\u6027&#xff1a;\u4e0d\u5962\u671b\u6a21\u578b\u5728\u4e00\u4e2aEP\u4e0a\u5b8c\u5168\u8fd0\u884c&#xff0c;\u53ef\u5728\u5f02\u6784\u8bbe\u5907\u4e0a\u534f\u540c\u5de5\u4f5c&#xff0c;\u5f3a\u8c03\u517c\u5bb9\u6027<\/li>\n<li>\u56fe\u8f6c\u6362&#xff1a;ORT\u7684\u6838\u5fc3\u529f\u80fd&#xff0c;\u4e3a\u6a21\u578b\u8ba1\u7b97\u56fe\u8f6c\u6362\u63d0\u4f9b\u4f18\u5316\u529f\u80fd&#xff1a;\u5168\u5c40\u4f18\u5316&#xff0c;\u5177\u4f53\u7b97\u5b50\u4f18\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3>EP<\/h3>\n<p>Execution Providers&#xff0c;\u6267\u884c\u63d0\u4f9b\u8005&#xff0c;\u901a\u8fc7\u63d2\u4ef6\u5316\u7684\u67b6\u6784&#xff0c;\u5141\u8bb8\u540c\u4e00\u4e2aONNX\u6a21\u578b\u5728\u4e0d\u540c\u786c\u4ef6\u4e0a\u65e0\u7f1d\u5207\u6362\u8fd0\u884c&#xff1a;<\/p>\n<table>\n<tr>\u6267\u884c\u63d0\u4f9b\u8005\u786c\u4ef6\u5e73\u53f0\u4f7f\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td align=\"left\">CPUExecutionProvider<\/td>\n<td align=\"left\">CPU<\/td>\n<td align=\"left\">\u901a\u7528\u63a8\u7406&#xff0c;\u65e0GPU\u73af\u5883<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">CUDAExecutionProvider<\/td>\n<td align=\"left\">NVIDIA GPU<\/td>\n<td align=\"left\">GPU\u52a0\u901f\u63a8\u7406<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">TensorrtExecutionProvider<\/td>\n<td align=\"left\">NVIDIA GPU<\/td>\n<td align=\"left\">\u901a\u8fc7TensorRT\u8fdb\u4e00\u6b65\u4f18\u5316<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">DirectMLExecutionProvider<\/td>\n<td align=\"left\">Windows GPU<\/td>\n<td align=\"left\">\u652f\u6301AMD\/Intel\/NVIDIA GPU<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">OpenVINOExecutionProvider<\/td>\n<td align=\"left\">Intel CPU\/GPU\/VPU<\/td>\n<td align=\"left\">Intel\u786c\u4ef6\u4f18\u5316<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">CoreMLExecutionProvider<\/td>\n<td align=\"left\">Apple Silicon<\/td>\n<td align=\"left\">macOS\/iOS\u4f18\u5316<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">XNNPACKExecutionProvider<\/td>\n<td align=\"left\">CPU\/\u79fb\u52a8\u7aef<\/td>\n<td align=\"left\">\u79fb\u52a8\u7aef\u548c\u8fb9\u7f18\u8bbe\u5907<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">WebGPUExecutionProvider<\/td>\n<td align=\"left\">\u6d4f\u89c8\u5668<\/td>\n<td align=\"left\">Web\u7aef\u63a8\u7406<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">ACLExecutionProvider<\/td>\n<td align=\"left\">ARM CPU<\/td>\n<td align=\"left\">ARM\u67b6\u6784\u4f18\u5316<\/td>\n<\/tr>\n<tr>\n<td align=\"left\">ROCmExecutionProvider<\/td>\n<td align=\"left\">AMD GPU<\/td>\n<td align=\"left\">AMD GPU\u52a0\u901f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u9009\u62e9\u4e0d\u540cEP<\/p>\n<p><span class=\"token keyword\">import<\/span> onnxruntime <span class=\"token keyword\">as<\/span> ort<\/p>\n<p><span class=\"token comment\"># CPU\u63a8\u7406<\/span><br \/>\nsession_cpu <span class=\"token operator\">&#061;<\/span> ort<span class=\"token punctuation\">.<\/span>InferenceSession<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#039;model.onnx&#039;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token comment\"># GPU\u63a8\u7406&#xff08;CUDA&#xff09;<\/span><br \/>\nsession_gpu <span class=\"token operator\">&#061;<\/span> ort<span class=\"token punctuation\">.<\/span>InferenceSession<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#039;model.onnx&#039;<\/span><span class=\"token punctuation\">,<\/span> providers<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;CUDAExecutionProvider&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;CPUExecutionProvider&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token comment\"># TensorRT\u63a8\u7406&#xff08;NVIDIA GPU\u4e0a\u6781\u81f4\u4f18\u5316&#xff09;<\/span><br \/>\nsession_trt <span class=\"token operator\">&#061;<\/span> ort<span class=\"token punctuation\">.<\/span>InferenceSession<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#039;model.onnx&#039;<\/span><span class=\"token punctuation\">,<\/span> providers<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;TensorrtExecutionProvider&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;CUDAExecutionProvider&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h3>ORT\u56fe\u4f18\u5316<\/h3>\n<p>\u56fe\u8f6c\u6362&#xff0c;\u5206\u4e3a\u56db\u4e2a\u5c42\u6b21&#xff0c;\u53c2\u8003core\/optimizer\/graph_transformer_level.h\u4ee3\u7801&#xff1a;<\/p>\n<p><span class=\"token keyword\">enum<\/span> <span class=\"token class-name\">class<\/span> TransformerLevel <span class=\"token operator\">:<\/span> <span class=\"token keyword\">int<\/span> <span class=\"token punctuation\">{<\/span><br \/>\nDefault <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\">\/\/ required transformers only<\/span><br \/>\nLevel1<span class=\"token punctuation\">,<\/span>       <span class=\"token comment\">\/\/ basic optimizations<\/span><br \/>\nLevel2<span class=\"token punctuation\">,<\/span>       <span class=\"token comment\">\/\/ extended optimizations<\/span><br \/>\nLevel3<span class=\"token punctuation\">,<\/span>       <span class=\"token comment\">\/\/ layout optimizations<\/span><br \/>\n<span class=\"token comment\">\/\/ The max level should always be same as the last level.<\/span><br \/>\nMaxLevel <span class=\"token operator\">&#061;<\/span> Level3<br \/>\n<span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">;<\/span><\/p>\n<p>\u89e3\u8bfb&#xff1a;<\/p>\n<ul>\n<li>Default\u5c42\u6b21&#xff1a;\u4e0d\u7ba1\u5982\u4f55\u914d\u7f6e&#xff0c;\u90fd\u4f1a\u8fdb\u884c\u65e0\u6761\u4ef6\u8f6c\u6362&#xff0c;\u5927\u90e8\u5206\u60c5\u51b5\u662f\u56fe\u8f6c\u6362\u524d\u5c31\u4f18\u5316&#xff0c;\u4e5f\u6709\u56fe\u8f6c\u6362\u540e\u518d\u505a\u4f18\u5316<\/li>\n<li>Level1&#xff1a;\u57fa\u7840\u4f18\u5316&#xff0c;\u548c\u540e\u7aefEP\u65e0\u5173&#xff0c;\u901a\u7528\u578b\u4f18\u5316&#xff0c;\u5982\u5e38\u91cf\u6298\u53e0\u3001\u7b97\u5b50\u878d\u5408<\/li>\n<li>Level2&#xff1a;\u6269\u5c55\u4f18\u5316&#xff0c;\u4e00\u822c\u9488\u5bf9\u7279\u5b9a\u540e\u7aefEP\u8fdb\u884c\u7b97\u5b50\u878d\u5408\u4f18\u5316<\/li>\n<li>Level3&#xff1a;Layout\u5e03\u5c40\u4f18\u5316&#xff0c;\u76ee\u524d\u5c31\u53ea\u6709\u4e00\u4e2aNhwcTransformer<\/li>\n<\/ul>\n<h2>\u62d3\u5c55<\/h2>\n<h3>Olive<\/h3>\n<p>GitHub\u3002<\/p>\n<p>\u529f\u80fd&#xff1a;\u5fae\u8c03\u3001\u91cf\u5316\u3002<\/p>\n<p>olive quantize\u5b50\u547d\u4ee4\u53ef\u652f\u6301\u5404\u79cd\u91cf\u5316\u65b9\u6cd5&#xff0c;\u5305\u62ecAWQ\u3001GPTQ\u3001BitsAndBytes\u3001ORT\u9759\u6001\u65b9\u6cd5\u3002<\/p>\n<p>\u91cf\u5316\u793a\u4f8b&#xff1a;<\/p>\n<p>olive quantize <span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> meta-llama\/Llama-3.2-1B-Instruct <span class=\"token parameter variable\">&#8211;algorithm<\/span> awq <span class=\"token parameter variable\">&#8211;output_path<\/span> models\/llama\/awq <span class=\"token parameter variable\">&#8211;log_level<\/span> <span class=\"token number\">1<\/span><br \/>\n<span class=\"token comment\"># \u5e26\u6821\u51c6\u6570\u636e<\/span><br \/>\nolive quantize <span class=\"token parameter variable\">&#8211;algorithm<\/span> gptq <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> meta-llama\/Llama-3.2-1B-Instruct  <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;data_name<\/span> wikitext <span class=\"token parameter variable\">&#8211;subset<\/span> wikitext-2-raw-v1 <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;split<\/span> train <span class=\"token parameter variable\">&#8211;max_samples<\/span> <span class=\"token number\">128<\/span> <span class=\"token parameter variable\">&#8211;output_path<\/span> models\/gptq<\/p>\n<p>\u4f18\u5316\u793a\u4f8b&#xff1a;<\/p>\n<p>olive auto-opt <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> models\/llama\/awq <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;device<\/span> cpu <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;provider<\/span> CPUExecutionProvider <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;use_ort_genai<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;output_path<\/span> models\/llama\/onnx <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;log_level<\/span> <span class=\"token number\">1<\/span><\/p>\n<p>\u5c06\u91cf\u5316\u6a21\u578b\u4f18\u5316\u4e3aONNX\u6a21\u578b\u3002<\/p>\n<p>\u5148\u91cf\u5316\u518d\u5fae\u8c03&#xff0c;\u91cf\u5316\u5e26\u6765\u7684\u635f\u5931\u53ef\u80fd\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u5f97\u4ee5\u6821\u51c6&#xff1a;<\/p>\n<p>olive quantize <span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> meta-llama\/Llama-3.2-1B-Instruct <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;trust_remote_code<\/span> <span class=\"token parameter variable\">&#8211;algorithm<\/span> awq <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;output_path<\/span> models\/llama\/awq <span class=\"token parameter variable\">&#8211;log_level<\/span> <span class=\"token number\">1<\/span><\/p>\n<p>olive finetune <span class=\"token parameter variable\">&#8211;method<\/span> lora <span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> models\/llama\/awq <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;data_name<\/span> xxyyzzz\/phrase_classification <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;text_template<\/span> <span class=\"token string\">&#034;&lt;|start_header_id|&gt;user&lt;|end_header_id|&gt;<span class=\"token entity\" title=\"\\\\n\">\\\\n<\/span>{phrase}&lt;|eot_id|&gt;&lt;|start_header_id|&gt;assistant&lt;|end_header_id|&gt;<span class=\"token entity\" title=\"\\\\n\">\\\\n<\/span>{tone}&#034;<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;max_steps<\/span> <span class=\"token number\">100<\/span> <span class=\"token parameter variable\">&#8211;output_path<\/span> .\/models\/llama\/ft <span class=\"token parameter variable\">&#8211;log_level<\/span> <span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># \u81ea\u52a8\u4f18\u5316<\/span><br \/>\nolive auto-opt <span class=\"token parameter variable\">&#8211;model_name_or_path<\/span> models\/llama\/ft\/model <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;adapter_path<\/span> models\/llama\/ft\/adapter <span class=\"token parameter variable\">&#8211;device<\/span> cpu <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;provider<\/span> CPUExecutionProvider <span class=\"token parameter variable\">&#8211;use_ort_genai<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;output_path<\/span> models\/llama\/onnx <span class=\"token parameter variable\">&#8211;log_level<\/span> <span class=\"token number\">1<\/span><\/p>\n<h3>\u53ef\u89c6\u5316<\/h3>\n<p>ONNX\u53ef\u89c6\u5316\u5de5\u5177\u6709\u5f88\u591a&#xff0c;\u53c2\u8003\u6a21\u578b\u53ef\u89c6\u5316\u9879\u76ee\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6982\u8ff0<br \/>\n\u5b98\u7f51&#xff0c;ONNX\u662fOpen Neural Network Exchange\u7684\u7f29\u5199&#xff0c;\u4f5c\u4e3a\u4e00\u79cd\u5f00\u653e\u5f00\u6e90&#xff08;GitHub&#xff0c;21.2K Star&#xff0c;4K Fork&#xff09;\u7684\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362\u683c\u5f0f&#xff0c;\u80fd\u591f\u5b9e\u73b0\u4e0d\u540c\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u95f4\u7684\u6a21\u578b\u8f6c\u6362\u4e0e\u5171\u4eab&#xff0c;\u4f7f\u5f97\u57fa\u4e8eONNX\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u591a\u79cd\u73af\u5883\u4e0b\u9ad8\u6548\u8fd0\u884c\u3002<br \/>\n\u5728LLM\u4e4b\u524d&#xff08;\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3&#xff09;\u5c31\u662f\u975e\u5e38\u6d41\u884c\u7684\u6a21\u578b\u683c\u5f0f&amp;#x<\/p>\n","protected":false},"author":2,"featured_media":84628,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9137,9477],"topic":[],"class_list":["post-84635","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-atomgit","tag-onnx"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \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\/84635.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6982\u8ff0 \u5b98\u7f51&#xff0c;ONNX\u662fOpen Neural Network Exchange\u7684\u7f29\u5199&#xff0c;\u4f5c\u4e3a\u4e00\u79cd\u5f00\u653e\u5f00\u6e90&#xff08;GitHub&#xff0c;21.2K Star&#xff0c;4K Fork&#xff09;\u7684\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362\u683c\u5f0f&#xff0c;\u80fd\u591f\u5b9e\u73b0\u4e0d\u540c\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u95f4\u7684\u6a21\u578b\u8f6c\u6362\u4e0e\u5171\u4eab&#xff0c;\u4f7f\u5f97\u57fa\u4e8eONNX\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u591a\u79cd\u73af\u5883\u4e0b\u9ad8\u6548\u8fd0\u884c\u3002 \u5728LLM\u4e4b\u524d&#xff08;\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3&#xff09;\u5c31\u662f\u975e\u5e38\u6d41\u884c\u7684\u6a21\u578b\u683c\u5f0f&amp;#x\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/84635.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-26T07:28:33+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072831-6a65b71fed5c7.png\" \/>\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=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/84635.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/84635.html\",\"name\":\"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-26T07:28:33+00:00\",\"dateModified\":\"2026-07-26T07:28:33+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/84635.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/84635.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/84635.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime\"}]},{\"@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":"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \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\/84635.html","og_locale":"zh_CN","og_type":"article","og_title":"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6982\u8ff0 \u5b98\u7f51&#xff0c;ONNX\u662fOpen Neural Network Exchange\u7684\u7f29\u5199&#xff0c;\u4f5c\u4e3a\u4e00\u79cd\u5f00\u653e\u5f00\u6e90&#xff08;GitHub&#xff0c;21.2K Star&#xff0c;4K Fork&#xff09;\u7684\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362\u683c\u5f0f&#xff0c;\u80fd\u591f\u5b9e\u73b0\u4e0d\u540c\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u95f4\u7684\u6a21\u578b\u8f6c\u6362\u4e0e\u5171\u4eab&#xff0c;\u4f7f\u5f97\u57fa\u4e8eONNX\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u591a\u79cd\u73af\u5883\u4e0b\u9ad8\u6548\u8fd0\u884c\u3002 \u5728LLM\u4e4b\u524d&#xff08;\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3&#xff09;\u5c31\u662f\u975e\u5e38\u6d41\u884c\u7684\u6a21\u578b\u683c\u5f0f&amp;#x","og_url":"https:\/\/www.wsisp.com\/helps\/84635.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-26T07:28:33+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260726072831-6a65b71fed5c7.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/84635.html","url":"https:\/\/www.wsisp.com\/helps\/84635.html","name":"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-26T07:28:33+00:00","dateModified":"2026-07-26T07:28:33+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/84635.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/84635.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/84635.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"ONNX\u751f\u6001\u7b80\u4ecb\u4e0e\u5b9e\u6218\uff1aONNX Runtime"}]},{"@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\/84635","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=84635"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/84635\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/84628"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=84635"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=84635"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=84635"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=84635"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}