{"id":16283,"date":"2024-07-17T16:56:54","date_gmt":"2024-07-17T08:56:54","guid":{"rendered":"http:\/\/139.9.1.231\/?p=16283"},"modified":"2024-07-17T16:56:56","modified_gmt":"2024-07-17T08:56:56","slug":"pytorchdistributeddataparallel","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/07\/17\/pytorchdistributeddataparallel\/","title":{"rendered":"pytorch\u591a\u673a\u591a\u5361\u8bad\u7ec3\u3010DistributedDataParallel\u3011"},"content":{"rendered":"\n<p><a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#distributeddataparallel\">https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#distributeddataparallel<\/a><\/p>\n\n\n\n\n\n<h2>\u539f\u7406<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"828\" height=\"669\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-95.png\" alt=\"\" class=\"wp-image-16287\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-95.png 828w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-95-300x242.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-95-768x621.png 768w\" sizes=\"(max-width: 828px) 100vw, 828px\" \/><\/figure>\n\n\n\n<p>DDP\u7684\u6d41\u7a0b\u793a\u610f\u56fe\u5982\u4e0a\u56fe\u6240\u793a\uff0cDDP\u9700\u8981\u989d\u5916\u7684\u5efa\u7acb\u8fdb\u7a0b\u7ec4\u9636\u6bb5\uff08Construction\uff09\u3002\u5728Construction\u9636\u6bb5\u9700\u8981\u9996\u5148\u660e\u786e\u901a\u4fe1\u534f\u8bae\u548c\u603b\u8fdb\u7a0b\u6570\u3002\u901a\u4fe1\u534f\u8bae\u662f\u5b9e\u73b0DDP\u7684\u5e95\u5c42\u57fa\u7840\uff0c\u6211\u4eec\u5728\u4e4b\u540e\u5355\u72ec\u4ecb\u7ecd\u3002\u603b\u8fdb\u7a0b\u6570\u5c31\u662f\u6307\u6709\u591a\u5c11\u4e2a\u72ec\u7acb\u7684\u5e76\u884c\u8fdb\u7a0b\uff0c\u88ab\u79f0\u4e3aworldsize\u3002\u6839\u636e\u9700\u6c42\u6bcf\u4e2a\u8fdb\u7a0b\u53ef\u4ee5\u5360\u7528\u4e00\u4e2a\u6216\u591a\u4e2aGPU\uff0c\u4f46\u5e76\u4e0d\u63a8\u8350\u591a\u4e2a\u8fdb\u7a0b\u5171\u4eab\u4e00\u4e2aGPU\uff0c\u8fd9\u4f1a\u9020\u6210\u6f5c\u5728\u7684\u6027\u80fd\u635f\u5931\u3002\u4e3a\u4e86\u4fbf\u4e8e\u7406\u89e3\uff0c\u5728\u672c\u6587\u7684\u6240\u6709\u793a\u4f8b\u4e2d\u6211\u4eec\u5047\u5b9a\u6bcf\u4e2a\u8fdb\u7a0b\u53ea\u5360\u75281\u4e2aGPU\uff0c\u5360\u7528\u591a\u4e2aGPU\u7684\u60c5\u51b5\u53ea\u9700\u8981\u7b80\u5355\u7684\u8c03\u6574GPU\u6620\u5c04\u5173\u7cfb\u5c31\u597d\u3002<\/p>\n\n\n\n<p>\u5e76\u884c\u7ec4\u5efa\u7acb\u4e4b\u540e\uff0c\u6bcf\u4e2aGPU\u4e0a\u4f1a\u72ec\u7acb\u7684\u6784\u5efa\u6a21\u578b\uff0c\u7136\u540eGPU-1\u4e2d\u6a21\u578b\u7684\u72b6\u6001\u4f1a\u88ab\u5e7f\u64ad\u5230\u5176\u5b83\u6240\u6709\u8fdb\u7a0b\u4e2d\u4ee5\u4fdd\u8bc1\u6240\u6709\u6a21\u578b\u90fd\u5177\u6709\u76f8\u540c\u7684\u521d\u59cb\u72b6\u6001\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662fConstruction\u53ea\u5728\u8bad\u7ec3\u5f00\u59cb\u524d\u6267\u884c\uff0c\u5728\u8bad\u7ec3\u4e2d\u53ea\u4f1a\u4e0d\u65ad\u8fed\u4ee3\u524d\u5411\u548c\u540e\u5411\u8fc7\u7a0b\uff0c\u56e0\u6b64\u4e0d\u4f1a\u5e26\u6765\u989d\u5916\u7684\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<p>\u76f8\u6bd4\u4e8e<code>DataParallel<\/code>\uff0cDDP\u7684\u524d\u5411\u540e\u5411\u8fc7\u7a0b\u66f4\u52a0\u7b80\u6d01\u3002\u63a8\u7406\u3001\u635f\u5931\u51fd\u6570\u8ba1\u7b97\uff0c\u68af\u5ea6\u8ba1\u7b97\u90fd\u662f\u5e76\u884c\u72ec\u7acb\u5b8c\u6210\u7684\u3002DDP\u5b9e\u73b0\u5e76\u884c\u8bad\u7ec3\u7684\u6838\u5fc3\u5728\u4e8e<strong>\u68af\u5ea6\u540c\u6b65<\/strong>\u3002\u68af\u5ea6\u5728\u6a21\u578b\u95f4\u7684\u540c\u6b65\u4f7f\u7528\u7684\u662f<code>allreduce<\/code>\u901a\u4fe1\u64cd\u4f5c\uff0c\u6bcf\u4e2aGPU\u4f1a\u5f97\u5230\u5b8c\u5168\u76f8\u540c\u7684\u68af\u5ea6\u3002\u5982\u56fe\u4e2d\u540e\u5411\u8fc7\u7a0b\u7684\u6b65\u9aa42\uff0cGPU\u95f4\u7684\u901a\u4fe1\u5728\u68af\u5ea6\u8ba1\u7b97\u5b8c\u6210\u540e\u88ab\u89e6\u53d1\uff08hook\u51fd\u6570\uff09\u3002\u56fe\u4e2d\u6ca1\u6709\u753b\u51fa\u7684\u662f\uff0c\u901a\u5e38\u6bcf\u4e2aGPU\u4e5f\u4f1a\u5efa\u7acb\u72ec\u7acb\u7684\u4f18\u5316\u5668\u3002\u7531\u4e8e\u6a21\u578b\u5177\u6709\u540c\u6837\u7684\u521d\u59cb\u72b6\u6001\u548c\u540e\u7eed\u76f8\u540c\u7684\u68af\u5ea6\uff0c\u56e0\u6b64\u6bcf\u8f6e\u8fed\u4ee3\u540e\u4e0d\u540c\u8fdb\u7a0b\u95f4\u7684\u6a21\u578b\u662f\u5b8c\u5168\u76f8\u540c\u7684\uff0c\u8fd9\u4fdd\u8bc1\u4e86DDP\u7684\u6570\u7406\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u4f18\u5316\u6027\u80fd\uff0cDDP\u4e2d\u9488\u5bf9<code>allreduce<\/code>\u64cd\u4f5c\u8fdb\u884c\u4e86\u66f4\u6df1\u5165\u7684\u8bbe\u8ba1\u3002\u68af\u5ea6\u7684\u8ba1\u7b97\u8fc7\u7a0b\u548c\u8fdb\u7a0b\u95f4\u7684\u901a\u4fe1\u8fc7\u7a0b\u5206\u522b\u9700\u8981\u6d88\u8017\u4e00\u5b9a\u91cf\u7684\u65f6\u95f4\u3002\u7b49\u5f85\u6a21\u578b\u6240\u6709\u7684\u53c2\u6570\u90fd\u8ba1\u7b97\u5b8c\u68af\u5ea6\u518d\u8fdb\u884c\u901a\u4fe1\u663e\u7136\u4e0d\u662f\u6700\u4f18\u7684\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cDDP\u4e2d\u7684\u8bbe\u8ba1\u662f\u901a\u8fc7\u5c06\u5168\u90e8\u6a21\u578b\u53c2\u6570\u5212\u5206\u4e3a\u65e0\u6570\u4e2a\u5c0f\u7684bucket\uff0c\u5728bucket\u7ea7\u522b\u5efa\u7acb<code>allreduce<\/code>\u3002\u5f53\u6240\u6709\u8fdb\u7a0b\u4e2dbucket0\u7684\u68af\u5ea6\u8ba1\u7b97\u5b8c\u6210\u540e\u5c31\u7acb\u523b\u5f00\u59cb\u901a\u4fe1\uff0c\u6b64\u65f6bucket1\u4e2d\u68af\u5ea6\u8fd8\u5728\u8ba1\u7b97\u3002\u8fd9\u6837\u53ef\u4ee5\u5b9e\u73b0\u8ba1\u7b97\u548c\u901a\u4fe1\u8fc7\u7a0b\u7684\u65f6\u95f4\u91cd\u53e0\u3002\u8fd9\u79cd\u8bbe\u8ba1\u80fd\u591f\u4f7f\u5f97DDP\u7684\u8bad\u7ec3\u66f4\u9ad8\u6548\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"822\" height=\"424\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-96.png\" alt=\"\" class=\"wp-image-16288\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-96.png 822w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-96-300x155.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-96-768x396.png 768w\" sizes=\"(max-width: 822px) 100vw, 822px\" \/><\/figure>\n\n\n\n<p>\u5728\u6700\u540e\u6211\u4eec\u5bf9DDP\u7684\u901a\u4fe1\u90e8\u5206\u8fdb\u884c\u4ecb\u7ecd\u3002DDP\u540e\u7aef\u7684\u901a\u4fe1\u7531\u591a\u79cdCPP\u7f16\u5199\u7684\u534f\u8bae\u652f\u6301\uff0c\u4e0d\u540c\u534f\u8bae\u5177\u6709\u4e0d\u540c\u7684\u901a\u4fe1\u7b97\u5b50\u7684\u652f\u6301\uff0c\u5728\u5f00\u53d1\u4e2d\u53ef\u4ee5\u6839\u636e\u9700\u6c42\u9009\u62e9\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-97.png\" alt=\"\" class=\"wp-image-16289\" width=\"440\" height=\"442\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-97.png 604w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-97-300x300.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-97-150x150.png 150w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/image-97-120x120.png 120w\" sizes=\"(max-width: 440px) 100vw, 440px\" \/><\/figure><\/div>\n\n\n\n<p>\u5bf9\u4e8eCV\u548cNLP\u5e38\u7528GPU\u8bad\u7ec3\u7684\u4efb\u52a1\u800c\u8a00\uff0c\u9009\u62e9Gloo\u6216NCCL\u534f\u8bae\u5373\u53ef\u3002\u4e00\u4e2a\u51b3\u5b9a\u56e0\u7d20\u662f\u4f60\u4f7f\u7528\u7684\u8ba1\u7b97\u673a\u96c6\u7fa4\u7684\u7f51\u7edc\u73af\u5883\uff1a<\/p>\n\n\n\n<ul><li><strong>\u5f53\u4f7f\u7528\u7684\u662fEthernet\uff08\u4ee5\u592a\u7f51\uff0c\u5927\u90e8\u5206\u673a\u5668\u90fd\u662f\u8fd9\u4e2a\u73af\u5883\uff09<\/strong>\uff1a\u90a3\u4e48\u4f18\u5148\u9009\u62e9NCCL\uff0c\u5177\u6709\u66f4\u597d\u7684\u6027\u80fd\uff1b\u5982\u679c\u5728\u4f7f\u7528\u4e2d\u9047\u5230\u4e86NCCL\u901a\u4fe1\u7684\u95ee\u9898\uff0c\u90a3\u4e48\u5c31\u9009\u62e9Gloo\u4f5c\u4e3a\u5907\u7528\u3002\uff08\u7ecf\u9a8c\uff1a\u5355\u673a\u591a\u5361\u76f4\u63a5NCCL\uff1b\u591a\u673a\u591a\u5361\u5148\u5c1d\u8bd5NCCL\uff0c\u5982\u679c\u901a\u4fe1\u6709\u95ee\u9898\uff0c\u800c\u4e14\u81ea\u5df1\u89e3\u51b3\u4e0d\u4e86\uff0c\u90a3\u5c31Gloo\u3002\uff09<\/li><li><strong>\u5f53\u4f7f\u7528\u7684\u662fInfiniBand<\/strong>\uff1a\u53ea\u652f\u6301NCCL\u3002<\/li><\/ul>\n\n\n\n<p>\u53e6\u4e00\u4e2a\u51b3\u5b9a\u6027\u56e0\u7d20\u662f\u4e8c\u8005\u652f\u6301\u7684\u7b97\u5b50\u8303\u56f4\u4e0d\u540c\uff0c\u56e0\u6b64\u5728\u4f7f\u7528\u65f6\u8fd8\u9700\u8981\u7ed3\u5408\u4ee3\u7801\u91cc\u7684\u529f\u80fd\u6765\u786e\u5b9a\u3002\u4e0b\u56fe\u8bb0\u5f55\u4e86\u6bcf\u79cd\u901a\u4fe1\u534f\u8bae\u80fd\u591f\u652f\u6301\u7684\u7b97\u5b50\uff0cGloo\u80fd\u591f\u5b9e\u73b0GPU\u4e2d\u6700\u57fa\u672c\u7684DDP\u8bad\u7ec3\uff0c\u800cNCCL\u80fd\u591f\u652f\u6301\u66f4\u52a0\u591a\u6837\u7684\u7b97\u5b50<\/p>\n\n\n\n<p>\u7efc\u4e0a\uff0c\u5f97\u76ca\u4e8eDDP\u7684\u5206\u5e03\u5f0f\u5e76\u884c\u8bbe\u8ba1\uff0cDDP\u5e76\u4e0d\u53d7PythonGIL\u4e89\u7528\u7684\u5f71\u54cd\uff0c\u662f\u4ee5\u591a\u8fdb\u7a0b\u7684\u65b9\u5f0f\u8fd0\u884c\u7684\u3002\u8fd9\u4e5f\u4f7f\u5f97DDP\u53ef\u4ee5\u652f\u6301\u591a\u673a\u591a\u5361\u7684\u8bad\u7ec3\u3002\u6211\u4eec\u5c06DDP\u7684\u4f18\u7f3a\u70b9\u6982\u62ec\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"979\" height=\"1024\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1-979x1024.png\" alt=\"\" class=\"wp-image-16292\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1-979x1024.png 979w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1-287x300.png 287w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1-768x803.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1-1469x1536.png 1469w, http:\/\/139.9.1.231\/wp-content\/uploads\/2024\/07\/Backends_Difference-1.png 1702w\" sizes=\"(max-width: 979px) 100vw, 979px\" \/><figcaption><em>\u4e0d\u540cBackend\u7684\u7b97\u5b50\u652f\u6301\u60c5\u51b5<\/em><\/figcaption><\/figure>\n\n\n\n<h3>\u4f18\u70b9<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E4%BC%98%E7%82%B9-1\"><\/a><\/h3>\n\n\n\n<ul><li>\u66f4\u5feb\u7684\u8bad\u7ec3\u901f\u5ea6<\/li><li>\u591a\u8fdb\u7a0b\u7684\u8fd0\u884c\u65b9\u5f0f<\/li><li>\u652f\u6301\u5355\u673a\u591a\u5361\u548c\u591a\u673a\u591a\u5361<\/li><li>\u5e73\u8861\u7684GPU\u4f7f\u7528<\/li><\/ul>\n\n\n\n<h3>\u7f3a\u70b9<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E7%BC%BA%E7%82%B9-1\"><\/a><\/h3>\n\n\n\n<ul><li>\u9700\u8981\u66f4\u591a\u7684\u4ee3\u7801\u4e66\u5199\u548c\u8bbe\u8ba1<\/li><\/ul>\n\n\n\n<h2>\u4ee3\u7801\u5b9e\u73b0\u548c\u53c2\u6570\u8bb2\u89e3\uff1a<\/h2>\n\n\n\n<p>\u672c\u6587\u9996\u5148\u4f1a\u57fa\u4e8eMNIST\u56fe\u50cf\u5206\u7c7b\u5efa\u7acb\u4e00\u4e2a\u6700\u5c0f\u539f\u578b\uff0c\u7136\u540e\u9010\u6b65\u6539\u8fdb\u5b83\u4ee5\u5b9e\u73b0\u591a\u673a\u591a\u5361\u7684\u8bad\u7ec3\u548c\u6df7\u5408\u7cbe\u5ea6\u7684\u652f\u6301\u3002\u5728\u8bb2\u8ff0\u7684\u601d\u8def\u4e0a\u672c\u6587\u501f\u9274\u4e86<a href=\"https:\/\/yangkky.github.io\/2019\/07\/08\/distributed-pytorch-tutorial.html\">Kevin Kaichuang Yang\u7684\u6559\u7a0b<\/a>\uff0c\u4f46\u5728\u5b9e\u73b0\u7ec6\u8282\u4e0a\u6709\u8f83\u5927\u7684\u5dee\u5f02\u3002\u7279\u522b\u7684\u662f\u672c\u6587\u589e\u52a0\u4e86\u5bf9DDP\u542f\u52a8\u65b9\u5f0f\u7684\u63a2\u8ba8\uff0c\u5e76\u4e14\u4ecb\u7ecd\u4e86\u591a\u8fdb\u7a0b\u901a\u4fe1\u64cd\u4f5c\u7684\u4f7f\u7528\u6837\u4f8b\u3002<\/p>\n\n\n\n<p>\u540d\u8bcd\u89e3\u91ca\uff1a\u4e00\u4e2aworld\u91cc\u8fdb\u7a0b\u4e2a\u6570\u4e3aworld_size\u3010\u5bf9\u4e8e2\u53612GPU\uff0c world_size =4\u3011\uff0c\u5168\u5c40\u770b\uff0c\u6bcf\u4e2a\u8fdb\u7a0b\u90fd\u6709\u4e00\u4e2a\u5e8f\u53f7rank\u30100\u4e3a\u4e3b\u673aGPU\u4e3b\u5361\u3011\uff1b\u5206\u5f00\u770b\uff0c\u4e00\u4e2a\u8fdb\u7a0b\u5728\u6bcf\u53f0\u673a\u5668\u91cc\u9762\u4e5f\u6709\u5e8f\u53f7local_rank\u3002<\/p>\n\n\n\n<ul><li>group\uff1a\u8fdb\u7a0b\u7ec4\uff0c\u9ed8\u8ba4\u4e00\u4e2a\u7ec4\uff0c\u5373\u4e00\u4e2aworld<\/li><li>world_size\uff1a\u5168\u5c40\u8fdb\u7a0b\u4e2a\u6570\u3010\u5bf9\u4e8e2\u53612GPU\uff0c world_size =4\u3011<\/li><li>rank\uff1a\u8fdb\u7a0b\u5e8f\u53f7\uff0c\u7528\u4e8e\u8fdb\u7a0b\u95f4\u901a\u4fe1\u3002rank=0\u4e3aGPU\u4e3b\u5361\uff0c\u4e3b\u8981\u7528\u4e8e\u591a\u673a\u591a\u5361\u3002\u672c\u6587\u4e2d\u4ec5\u6d89\u53ca\u5230\u4e00\u53f0\u673a\u5668\u5185\u591a\u5f20\u5361\u3002<\/li><li>locak_rank\uff1a\u8fdb\u7a0b\u5185\u7684GPU\u7f16\u53f7\uff0c\u901a\u8fc7\u6307\u4ee4<code>torch.distributed.run<\/code>\u81ea\u52a8\u6307\u5b9a\uff0c\u4e0d\u9700\u8981\u8ba4\u4e3a\u8bbe\u7f6e\u3002<\/li><\/ul>\n\n\n\n<h3>\u975e\u591a\u8fdb\u7a0b\u793a\u4f8b<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E9%9D%9E%E5%A4%9A%E8%BF%9B%E7%A8%8B%E7%A4%BA%E4%BE%8B\"><\/a><\/h3>\n\n\n\n<p>\u9996\u5148\u5f15\u5165\u4e86\u6240\u6709\u7528\u5230\u7684\u5e93\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">from datetime import datetime\nimport argparse\nimport torchvision\nimport torchvision.transforms as transforms\nimport torch\nimport torch.nn as nn\nimport torch.distributed as dist\nfrom tqdm import tqdm<\/pre>\n\n\n\n<p>\u5b9a\u4e49\u4e00\u4e2a\u7b80\u5355\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">class ConvNet(nn.Module):\n    def __init__(self, num_classes=10):\n        super(ConvNet, self).__init__()\n        self.layer1 = nn.Sequential(\n            nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2),\n            nn.BatchNorm2d(16),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.layer2 = nn.Sequential(\n            nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),\n            nn.BatchNorm2d(32),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.fc = nn.Linear(7*7*32, num_classes)\n\n    def forward(self, x):\n        out = self.layer1(x)\n        out = self.layer2(out)\n        out = out.reshape(out.size(0), -1)\n        out = self.fc(out)\n        return out<\/pre>\n\n\n\n<p>\u5b9a\u4e49\u4e3b\u51fd\u6570\uff0c\u6dfb\u52a0\u4e00\u4e9b\u542f\u52a8\u811a\u672c\u7684\u53ef\u9009\u53c2\u6570\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">def main():\n    parser = argparse.ArgumentParser()\n    parser.add_argument('-g', '--gpuid', default=0, type=int,\n                        help=\"which gpu to use\")\n    parser.add_argument('-e', '--epochs', default=2, type=int, \n                        metavar='N',\n                        help='number of total epochs to run')\n    parser.add_argument('-b', '--batch_size', default=4, type=int, \n                        metavar='N',\n                        help='number of batchsize')         \n\n    args = parser.parse_args()\n    train(args.gpuid, args)<\/pre>\n\n\n\n<p>\u7136\u540e\u7ed9\u51fa\u8bad\u7ec3\u51fd\u6570\u7684\u8be6\u7ec6\u5185\u5bb9\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">def train(gpu, args):\n    model = ConvNet()\n    model.cuda(gpu)\n    # define loss function (criterion) and optimizer\n    criterion = nn.CrossEntropyLoss().to(gpu)\n    optimizer = torch.optim.SGD(model.parameters(), 1e-4)\n\n    # Data loading code\n    train_dataset = torchvision.datasets.MNIST(root='.\/data',\n                                               train=True,\n                                               transform=transforms.ToTensor(),\n                                               download=True)\n    train_loader = torch.utils.data.DataLoader(dataset=train_dataset,\n                                               batch_size=args.batch_size,\n                                               shuffle=True,\n                                               num_workers=0,\n                                               pin_memory=True,\n                                               sampler=None)\n\n    start = datetime.now()\n    total_step = len(train_loader)\n    for epoch in range(args.epochs):\n        model.train()\n        for i, (images, labels) in enumerate(tqdm(train_loader)):\n            images = images.to(gpu)\n            labels = labels.to(gpu)\n            # Forward pass\n            outputs = model(images)\n            loss = criterion(outputs, labels)\n\n            # Backward and optimize\n            optimizer.zero_grad()\n            loss.backward()\n            optimizer.step()\n            if (i + 1) % 100 == 0:\n                print('Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}'.format(epoch + 1, args.epochs, i + 1, total_step,\n                                                                   loss.item()))\n    print(\"Training complete in: \" + str(datetime.now() - start))<\/pre>\n\n\n\n<p>\u6700\u540e\u786e\u4fdd\u4e3b\u51fd\u6570\u88ab\u542f\u52a8\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">if __name__ == '__main__':\n    main()<\/pre>\n\n\n\n<p>\u4ee5\u4e0a\u662f\u6211\u4eec\u7684MNIST\u56fe\u50cf\u5206\u7c7b\u6700\u5c0f\u539f\u578b\uff0c\u53ef\u4ee5\u901a\u8fc7\u5982\u4e0b\u547d\u4ee4\u542f\u52a8\u5728\u6307\u5b9a\u5355\u4e2aGPU\u4e0a\u7684\u8bad\u7ec3\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">python train.py -g 0<\/pre>\n\n\n\n<h3>\u591a\u8fdb\u7a0b\u793a\u4f8b<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E5%A4%9A%E8%BF%9B%E7%A8%8B%E7%A4%BA%E4%BE%8B\"><\/a><\/h3>\n\n\n\n<p>\u5728\u5f00\u59cb\u5bf9\u6700\u5c0f\u539f\u578b\u7684\u6539\u9020\u4e4b\u524d\uff0c\u6211\u4eec\u8fd8\u9700\u8981\u4ea4\u4ee3\u4e00\u4e9b\u4e8b\u60c5\u3002\u5728DDP\u7684\u4ee3\u7801\u5b9e\u73b0\u4e2d\uff0c\u6700\u91cd\u8981\u7684\u6b65\u9aa4\u4e4b\u4e00\u5c31\u662f\u521d\u59cb\u5316\u3002\u6240\u8c13\u521d\u59cb\u5316\u5bf9\u5e94\u4e8e\u4e0a\u6587\u4ecb\u7ecd\u7684Construction\u9636\u6bb5\uff0c\u6bcf\u4e2a\u8fdb\u7a0b\u4e2d\u9700\u8981\u6307\u660e\u51e0\u4e2a\u5173\u952e\u7684\u53c2\u6570\uff1a<\/p>\n\n\n\n<ul class=\"has-light-gray-background-color has-background\"><li><strong>backend<\/strong>\uff1a\u660e\u786e\u540e\u7aef\u901a\u4fe1\u65b9\u5f0f\uff0cNCCL\u8fd8\u662fGloo<\/li><li><strong>init_method<\/strong>\uff1a\u521d\u59cb\u5316\u65b9\u5f0f\uff0cTCP\u8fd8\u662fEnvironment variable\uff08Env\uff09\uff0c\u53ef\u4ee5\u7b80\u5355\u7406\u89e3\u4e3a\u8fdb\u7a0b\u83b7\u53d6\u5173\u952e\u53c2\u6570\u7684\u5730\u5740\u548c\u65b9\u5f0f<\/li><li><strong>world_size<\/strong>\uff1a\u603b\u7684\u8fdb\u7a0b\u6570\u6709\u591a\u5c11<\/li><li><strong>rank<\/strong>\uff1a\u5f53\u524d\u8fdb\u7a0b\u662f\u603b\u8fdb\u7a0b\u4e2d\u7684\u7b2c\u51e0\u4e2a<\/li><\/ul>\n\n\n\n<p>\u521d\u59cb\u5316\u65b9\u5f0f\u4e0d\u540c\u4f1a\u5f71\u54cd\u4ee3\u7801\u7684\u542f\u52a8\u90e8\u5206\u3002\u672c\u6587\u4f1a\u5206\u522b\u7ed9\u51faTCP\u548cENV\u6a21\u5f0f\u7684\u6837\u4f8b\u3002TCP\u6a21\u5f0f<\/p>\n\n\n\n<p>\u8ba9\u6211\u4eec\u5148\u4eceTCP\u5f00\u59cb\uff0c\u6ce8\u610f\u90a3\u4e9b\u6807\u8bb0\u88ab\u66f4\u6539\u7684\u4ee3\u7801\u90e8\u5206\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">def main():\n    parser = argparse.ArgumentParser()\n    parser.add_argument('-g', '--gpuid', default=0, type=int,\n                        help=\"which gpu to use\")\n    parser.add_argument('-e', '--epochs', default=1, type=int, \n                        metavar='N',\n                        help='number of total epochs to run')\n    parser.add_argument('-b', '--batch_size', default=4, type=int, \n                        metavar='N',\n                        help='number of batchsize')   \n    ##################################################################################\n    parser.add_argument('--init_method', default='tcp:\/\/localhost:18888',            #\n                        help=\"init-method\")                                          #\n    parser.add_argument('-r', '--rank', default=0, type=int,                         #\n                    help='rank of current process')                                  #\n    parser.add_argument('--world_size', default=2, type=int,                         #\n                        help=\"world size\")                                           #\n    parser.add_argument('--use_mix_precision', default=False,                        #\n                        action='store_true', help=\"whether to use mix precision\")    #\n    ##################################################################################                  \n    args = parser.parse_args()\n    train(args.gpuid, args)<\/pre>\n\n\n\n<p>\u5728main\u51fd\u6570\u4e2d\u9700\u8981\u589e\u52a0\u4e86\u4ee5\u4e0b\u53c2\u6570\uff1a<\/p>\n\n\n\n<ul class=\"has-light-gray-background-color has-background\"><li>args.init_method\uff1aurl\u5730\u5740\uff0c\u7528\u6765\u6307\u660e\u7684\u521d\u59cb\u5316\u65b9\u6cd5\u3002\u5728tcp\u521d\u59cb\u5316\u65b9\u6cd5\u4e2d\uff0c\u5176\u683c\u5f0f\u5e94\u4e3a\uff1atcp:[ IP ]:[ Port ] \u3002IP\u4e3arank=0\u8fdb\u7a0b\u6240\u5728\u7684\u673a\u5668IP\u5730\u5740\uff0cPort\u4e3a\u4efb\u610f\u4e00\u4e2a\u7a7a\u95f2\u7684\u7aef\u53e3\u53f7\u3002\u5f53<strong>\u91c7\u7528\u7684\u662f\u5355\u673a\u591a\u5361\u6a21\u5f0f\u65f6\uff0cIP\u53ef\u4ee5\u9ed8\u8ba4\u4e3a\/\/localhost<\/strong><\/li><li>args.rank\uff1a\u5f53\u524d\u8fdb\u7a0b\u5728\u6240\u6709\u8fdb\u7a0b\u4e2d\u7684\u5e8f\u53f7<\/li><li>args.world_size\uff1a\u8fdb\u7a0b\u603b\u6570\u3010\u4e00\u5171\u51e0\u5757GPU\u3011<\/li><li>args.use_mix_precision\uff1a\u5e03\u5c14\u53d8\u91cf\uff0c\u63a7\u5236\u662f\u5426\u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6<\/li><\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">def train(gpu, args):\n    ########################################    N1    ####################################################################\n    dist.init_process_group(backend='nccl', init_method=args.init_method, rank=args.rank, world_size=args.world_size)    #\n    ######################################################################################################################\n    model = ConvNet()\n    model.cuda(gpu)\n    # define loss function (criterion) and optimizer\n    criterion = nn.CrossEntropyLoss().to(gpu)\n    optimizer = torch.optim.SGD(model.parameters(), 1e-4)\n    # Wrap the model\n    #######################################    N2    ########################\n    model = nn.SyncBatchNorm.convert_sync_batchnorm(model)                  #\n    model = nn.parallel.DistributedDataParallel(model, device_ids=[gpu])    #\n    scaler = GradScaler(enabled=args.use_mix_precision)                   #\n    #########################################################################\n    # Data loading code\n    train_dataset = torchvision.datasets.MNIST(root='.\/data',\n                                               train=True,\n                                               transform=transforms.ToTensor(),\n                                               download=True)\n    ####################################    N3    #######################################\n    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)      #\n    train_loader = torch.utils.data.DataLoader(dataset=train_dataset,                   #\n                                               batch_size=args.batch_size,              #\n                                               shuffle=False,                           #\n                                               num_workers=0,                           #\n                                               pin_memory=True,                         #\n                                               sampler=train_sampler)                   #\n    #####################################################################################\n    start = datetime.now()\n    total_step = len(train_loader) # The number changes to orignal_length \/\/ args.world_size\n    for epoch in range(args.epochs):\n        ################    N4    ################\n        train_loader.sampler.set_epoch(epoch)    #\n        ##########################################\n        model.train()\n        for i, (images, labels) in enumerate(tqdm(train_loader)):\n            images = images.to(gpu)\n            labels = labels.to(gpu)\n            # Forward pass\n            ########################    N5    ################################\n            with torch.cuda.amp.autocast(enabled=args.use_mix_precision):    #\n                outputs = model(images)                                      #\n                loss = criterion(outputs, labels)                            #\n            ##################################################################  \n            # Backward and optimize\n            optimizer.zero_grad()\n            ##############    N6    ##########\n            scaler.scale(loss).backward()    #\n            scaler.step(optimizer)           #\n            scaler.update()                  #\n            ##################################\n            ################    N7    ####################\n            if (i + 1) % 100 == 0 and args.rank == 0:    #\n            ##############################################   \n                print('Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}'.format(epoch + 1, args.epochs, i + 1, total_step,\n                                                                   loss.item()))            \n    ############    N8    ###########\n    dist.destroy_process_group()    #                                       \n    if args.rank == 0:              #\n    #################################\n        print(\"Training complete in: \" + str(datetime.now() - start))<\/pre>\n\n\n\n<p>\u5728\u8bad\u7ec3\u51fd\u6570\u4e2d\u589e\u52a0\/\u4fee\u6539\u4e86\u4ee5\u4e0b\u5185\u5bb9\uff1a<\/p>\n\n\n\n<ul><li>N1\uff1a\u589e\u52a0\u4e86DDP\u521d\u59cb\u5316\u7684\u4ee3\u7801\uff0c\u9700\u8981\u6307\u660ebackend\u3001init_method\u3001rank\u548cworld_size\u3002\u5176\u542b\u4e49\u5728\u524d\u6587\u90fd\u6709\u4ecb\u7ecd\u3002<\/li><li>N2\uff1a\u5728\u5e76\u884c\u73af\u5883\u4e0b\uff0c\u5bf9\u4e8e\u7528\u5230BN\u5c42\u7684\u6a21\u578b\u9700\u8981\u8f6c\u6362\u4e3a\u540c\u6b65BN\u5c42\uff1b\u5176\u6b21\uff0c\u7528<code>DistributedDataParallel<\/code>\u5c06\u6a21\u578b\u5c01\u88c5\u4e3a\u4e00\u4e2aDDP\u6a21\u578b\uff0c\u5e76\u590d\u5236\u5230\u6307\u5b9a\u7684GPU\u4e0a\u3002\u5c01\u88c5\u65f6\u4e0d\u9700\u8981\u66f4\u6539\u6a21\u578b\u5185\u90e8\u7684\u4ee3\u7801\uff1b\u8bbe\u7f6e\u6df7\u5408\u7cbe\u5ea6\u4e2d\u7684scaler\uff0c\u901a\u8fc7\u8bbe\u7f6e<code>enabled<\/code>\u53c2\u6570\u63a7\u5236\u662f\u5426\u751f\u6548\u3002<\/li><li>N3\uff1aDDP\u8981\u6c42\u5b9a\u4e49<code>distributed.DistributedSampler<\/code>\uff0c\u901a\u8fc7\u5c01\u88c5<code>train_dataset<\/code>\u5b9e\u73b0\uff1b\u5728\u5efa\u7acb<code>DataLoader<\/code>\u65f6\u6307\u5b9a<code>sampler<\/code>\u3002\u6b64\u5916\u8fd8\u8981\u6ce8\u610f\uff1a<code>shuffle=False<\/code>\u3002DDP\u7684\u6570\u636e\u6253\u4e71\u9700\u8981\u901a\u8fc7\u8bbe\u7f6e<code>sampler<\/code>\uff0c\u53c2\u8003N4\u3002<\/li><li>N4\uff1a\u5728\u6bcf\u4e2aepoch\u5f00\u59cb\u524d\u6253\u4e71\u6570\u636e\u987a\u5e8f\u3002\uff08\u6ce8\u610ftotal_step\u5df2\u7ecf\u53d8\u4e3a<code>orignal_length \/\/ args.world_size<\/code>\u3002\uff09<\/li><li>N5\uff1a\u5229\u7528<code>torch.cuda.amp.autocast<\/code>\u63a7\u5236\u524d\u5411\u8fc7\u7a0b\u4e2d\u662f\u5426\u4f7f\u7528\u534a\u7cbe\u5ea6\u8ba1\u7b97\u3002<\/li><li>N6: \u5f53\u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6\u65f6\uff0cscaler\u4f1a\u7f29\u653eloss\u6765\u907f\u514d\u7531\u4e8e\u7cbe\u5ea6\u53d8\u5316\u5bfc\u81f4\u68af\u5ea6\u4e3a0\u7684\u60c5\u51b5\u3002<\/li><li>N7\uff1a\u4e3a\u4e86\u907f\u514dlog\u4fe1\u606f\u7684\u91cd\u590d\u6253\u5370\uff0c\u53ef\u4ee5\u53ea\u5141\u8bb8rank0\u53f7\u8fdb\u7a0b\u6253\u5370\u3002<\/li><li>N8: \u6e05\u7406\u8fdb\u7a0b\uff1b\u7136\u540e\uff0c\u540c\u4e0a\u3002<\/li><\/ul>\n\n\n\n<p>\u5047\u8bbe\u670d\u52a1\u5668\u73af\u5883\u4e3a2\u53f0\u670d\u52a1\u5668\uff08\u4e5f\u79f0\u4e3a2\u4e2anode\uff09\uff0c\u6bcf\u53f0\u670d\u52a1\u5668\u4e24\u5757GPU\u3002\u542f\u52a8\u65b9\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\"># Node 0 : ip 192.168.1.201  port : 12345\n# terminal-0\npython mnist-tcp.py --init_method tcp:\/\/192.168.1.201:12345 -g 0 --rank 0 --world_size 4 --use_mix_precision\n# terminal-1\npython mnist-tcp.py --init_method tcp:\/\/192.168.1.201:12345 -g 1 --rank 1 --world_size 4 --use_mix_precision\n\n# Node 1 : \n# terminal-0\npython tcp_init.py --init_method tcp:\/\/192.168.1.201:12345 -g 0 --rank 2 --world_size 4 --use_mix_precision\n# terminal-1\npython tcp_init.py --init_method tcp:\/\/192.168.1.201:12345 -g 1 --rank 3 --world_size 4 --use_mix_precision<\/pre>\n\n\n\n<p>TCP\u6a21\u5f0f\u542f\u52a8\u5f88\u597d\u7406\u89e3\uff0c\u9700\u8981\u5728bash\u4e2d\u72ec\u7acb\u7684\u542f\u52a8\u6bcf\u4e00\u4e2a\u8fdb\u7a0b\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u8fdb\u7a0b\u5206\u914d\u597d\u5176rank\u5e8f\u53f7\u3002\u7f3a\u70b9\u662f\u5f53\u8fdb\u7a0b\u6570\u591a\u7684\u65f6\u5019\u542f\u52a8\u6bd4\u8f83\u9ebb\u70e6\u3002\u5b8c\u6574\u7684\u811a\u672c\u6587\u4ef6\u89c1<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\">\u8fd9\u91cc<\/a>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<p>ENV\u6a21\u5f0f<\/p>\n\n\n\n<p>ENV\u6a21\u5f0f\u542f\u52a8\u4f1a\u66f4\u7b80\u6d01\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u8fdb\u7a0b\u5e76\u4e0d\u9700\u8981\u5728<code>dist.init_process_group<\/code>\u4e2d\u624b\u52a8\u7684\u6307\u5b9a\u5176rank\u3001world_size\u548curl\u3002\u7a0b\u5e8f\u4f1a\u5728\u73af\u5883\u53d8\u91cf\u4e2d\u53bb\u5bfb\u627e\u8fd9\u4e9b\u503c\u3002\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">def main():\n    parser = argparse.ArgumentParser()\n    parser.add_argument('-g', '--gpuid', default=0, type=int,\n                        help=\"which gpu to use\")\n    parser.add_argument('-e', '--epochs', default=1, type=int, \n                        metavar='N',\n                        help='number of total epochs to run')\n    parser.add_argument('-b', '--batch_size', default=4, type=int, \n                        metavar='N',\n                        help='number of batchsize')   \n    ##################################################################################\n    parser.add_argument(\"--local_rank\", type=int,                                    #\n                        help='rank in current node')                                 #\n    parser.add_argument('--use_mix_precision', default=False,                        #\n                        action='store_true', help=\"whether to use mix precision\")    #\n    ##################################################################################                  \n    args = parser.parse_args()\n    #################################\n    train(args.local_rank, args)    #\n    #################################<\/pre>\n\n\n\n<ul><li>args.local_rank\uff1a\u8fd9\u91cc\u6307\u7684\u662f\u5f53\u524d\u8fdb\u7a0b\u5728\u5f53\u524d\u673a\u5668\u4e2d\u7684\u5e8f\u53f7\uff0c\u6ce8\u610f\u548c\u5728\u5168\u90e8\u8fdb\u7a0b\u4e2d\u5e8f\u53f7\u7684\u533a\u522b\u3002\u5728ENV\u6a21\u5f0f\u4e2d\uff0c\u8fd9\u4e2a\u53c2\u6570\u662f\u5fc5\u987b\u7684\uff0c\u7531\u542f\u52a8\u811a\u672c\u81ea\u52a8\u5212\u5206\uff0c\u4e0d\u9700\u8981\u624b\u52a8\u6307\u5b9a\u3002\u8981\u5584\u7528<code>local_rank<\/code>\u6765\u5206\u914dGPU_ID\u3002<\/li><li><code>train(args.local_rank, args)<\/code>\uff1a\u4e00\u822c\u60c5\u51b5\u4e0b\u4fdd\u6301local_rank\u4e0e\u8fdb\u7a0b\u6240\u7528GPU_ID\u4e00\u81f4\u3002<\/li><\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">def train(gpu, args):\n    ##################################################################\n    dist.init_process_group(backend='nccl', init_method='env:\/\/')    #\n    args.rank = dist.get_rank()                                      #\n    ##################################################################\n    model = ConvNet()\n    ...<\/pre>\n\n\n\n<ul><li>\u8bad\u7ec3\u51fd\u6570\u4e2d\u4ec5\u9700\u8981\u66f4\u6539\u521d\u59cb\u5316\u65b9\u5f0f\u5373\u53ef\u3002\u5728ENV\u4e2d\u53ea\u9700\u8981\u6307\u5b9a<code>init_method='env:\/\/'<\/code>\u3002TCP\u6240\u9700\u7684\u5173\u952e\u53c2\u6570\u6a21\u578b\u4f1a\u4ece\u73af\u5883\u53d8\u91cf\u4e2d\u81ea\u52a8\u83b7\u53d6\uff0c\u73af\u5883\u53d8\u91cf\u53ef\u4ee5\u5728\u7a0b\u5e8f\u5916\u90e8\u542f\u52a8\u65f6\u8bbe\u5b9a\uff0c\u53c2\u8003\u542f\u52a8\u65b9\u5f0f\u3002<\/li><li>\u5f53\u524d\u8fdb\u7a0b\u7684rank\u503c\u53ef\u4ee5\u901a\u8fc7<code>dist.get_rank()<\/code>\u5f97\u5230<\/li><li>\u4e4b\u540e\u7684\u4ee3\u7801\u4e0eTCP\u5b8c\u5168\u76f8\u540c<\/li><\/ul>\n\n\n\n<p>\u5047\u8bbe\u670d\u52a1\u5668\u73af\u5883\u4e3a2\u53f0\u670d\u52a1\u5668\uff08\u4e5f\u79f0\u4e3a2\u4e2anode\uff09\uff0c\u6bcf\u53f0\u670d\u52a1\u5668\u4e24\u5757GPU\u3002ENV\u6a21\u5f0f\u7684\u542f\u52a8\u65b9\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\"># Node 0 : ip 192.168.1.201  port : 12345\n# terminal-0\npython -m torch.distributed.launch --nproc_per_node=2 --nnodes=2 --node_rank=0 --master_addr=\"192.168.1.201\" --master_port=12345 mnist-env.py --use_mix_precision\n\n# Node 1 : \n# terminal-0\npython -m torch.distributed.launch --nproc_per_node=2 --nnodes=2 --node_rank=1 --master_addr=\"192.168.1.201\" --master_port=12345 mnist-env.py --use_mix_precision<\/pre>\n\n\n\n<p>ENV\u6a21\u5f0f\u53ef\u4ee5\u4f7f\u7528pytorch\u4e2d\u7684\u542f\u52a8\u811a\u672c<code>torch.distributed.launch<\/code>\u542f\u52a8\u3002\u5728\u542f\u52a8\u547d\u4ee4\u4e2d\u9700\u8981\u6307\u660e\u591a\u4e2a\u53c2\u6570\uff1a<\/p>\n\n\n\n<ul class=\"has-light-gray-background-color has-background\"><li>nproc_per_node: \u6bcf\u53f0\u673a\u5668\u4e2d\u8fd0\u884c\u51e0\u4e2a\u8fdb\u7a0b\u3010\u6bcf\u53f0\u673a\u5668\u51e0\u4e2aGPU\u3011<\/li><li>nnodes\uff1a\u4e00\u5171\u4f7f\u7528\u591a\u5c11\u53f0\u673a\u5668<\/li><li>node_rank\uff1a\u5f53\u524d\u673a\u5668\u7684\u5e8f\u53f7\u3010\u975eGPU\u5e8f\u53f7\u3011<\/li><li>master_addr\uff1a0\u53f7\u673a\u5668\u7684IP<\/li><li>master_port\uff1a0\u53f7\u673a\u5668\u7684\u53ef\u7528\u7aef\u53e3<\/li><\/ul>\n\n\n\n<p>\u53ef\u4ee5\u770b\u5230\u65e0\u8bba\u4e00\u53f0\u673a\u5668\u4e2d\u7684\u8fdb\u7a0b\u6570\u4e3a\u591a\u5c11\uff0c\u53ea\u9700\u8981\u4e00\u884c\u547d\u4ee4\u5c31\u53ef\u4ee5\u542f\u52a8\uff0c\u76f8\u6bd4\u4e8eTCP\u6a21\u5f0f\u542f\u52a8\u65b9\u5f0f\u66f4\u52a0\u7b80\u6d01\u3002<\/p>\n\n\n\n<p>\u8bad\u7ec3\u4e2d\u5bf9\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\u8fdb\u884c\u9a8c\u8bc1\u4e5f\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u6b65\u9aa4\u4e4b\u4e00\uff0c\u90a3\u4e48\u5982\u4f55\u5728\u4e0a\u8ff0demo\u4e2d\u589e\u52a0\u6a21\u578b\u9a8c\u8bc1\u7684\u4ee3\u7801\u5462\uff1f\u5982\u4f55\u5b9e\u73b0\u6a21\u578b\u7684\u5e76\u884c\u9a8c\u8bc1\uff1f<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">####################################    N11    ##################################\ndef evaluate(model, gpu, test_loader, rank):\n    model.eval()\n    size = torch.tensor(0.).to(gpu)\n    correct = torch.tensor(0.).to(gpu)\n    with torch.no_grad():\n        for i, (images, labels) in enumerate(tqdm(test_loader)):\n            images = images.to(gpu)\n            labels = labels.to(gpu)\n            outputs = model(images)\n            size += images.shape[0]\n            correct += (outputs.argmax(1) == labels).type(torch.float).sum() \n    dist.reduce(size, 0, op=dist.ReduceOp.SUM) # \u7fa4\u4f53\u901a\u4fe1 reduce \u64cd\u4f5c change to allreduce if Gloo\n    dist.reduce(correct, 0, op=dist.ReduceOp.SUM) # \u7fa4\u4f53\u901a\u4fe1 reduce \u64cd\u4f5c change to allreduce if Gloo\n    if rank==0:\n        print('Evaluate accuracy is {:.2f}'.format(correct \/ size))\n #################################################################################\n\ndef train(gpu, args):\n    ...\n    ####################################    N9    ###################################\n    test_dataset = torchvision.datasets.MNIST(root='.\/data',                        #\n                                               train=False,                         #\n                                               transform=transforms.ToTensor(),     #\n                                               download=True)                       #\n    test_sampler = torch.utils.data.distributed.DistributedSampler(test_dataset)    #\n    test_loader = torch.utils.data.DataLoader(dataset=test_dataset,                 #\n                                               batch_size=args.batch_size,               #\n                                               shuffle=False,                       #\n                                               num_workers=0,                       #\n                                               pin_memory=True,                     #\n                                               sampler=test_sampler)                #\n    #################################################################################\n    start = datetime.now()\n    total_step = len(train_loader) # The number changes to orignal_length \/\/ args.world_size\n    for epoch in range(args.epochs):\n        ...\n        #####################    N10    #################\n        evaluate(model, gpu, test_loader, args.rank)    #\n        #################################################\n    ...        <\/pre>\n\n\n\n<p>\u7701\u7565\u4e86\u4ee3\u7801\u4e0d\u53d8\u7684\u90e8\u5206\uff0c\u5b8c\u6574\u7684\u7a0b\u5e8f\u89c1<a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\">\u811a\u672c<\/a>\u3002<\/p>\n\n\n\n<ul><li>N9\uff1a\u589e\u52a0\u9a8c\u8bc1\u96c6\u7684DataLoader\uff0c\u8bbe\u7f6esampler\u5b9e\u73b0\u6570\u636e\u7684\u5e76\u884c\u5207\u5206<\/li><li>N10\uff1a\u5728\u6bcf\u4e2aepoch\u7ed3\u675f\u524d\u9a8c\u8bc1\u6a21\u578b<\/li><li>N11: \u5229\u7528\u7fa4\u4f53\u901a\u4fe1<code>Reduce<\/code>\u64cd\u4f5c\uff0c\u5c06\u8ba1\u7b97\u51c6\u786e\u7387\u6240\u9700\u7684\u6b63\u786e\u9884\u6d4b\u6570\u548c\u5168\u5c40\u6837\u672c\u6570\u6536\u96c6\u5230rank0\u8fdb\u7a0b\u4e2d<\/li><\/ul>\n\n\n\n<p>\u53ea\u9700\u8981\u5229\u7528\u7fa4\u4f53\u901a\u4fe1\u5c06\u9a8c\u8bc1\u96c6\u6837\u672c\u6570\u548c\u9884\u6d4b\u6b63\u786e\u7684\u6837\u672c\u6570\u6c47\u96c6\u5728rank0\u4e2d\u5373\u53ef\u5b9e\u73b0\u5e76\u884c\u7684\u6a21\u578b\u9a8c\u8bc1\uff0c\u5bf9\u4e8e\u5176\u5b83\u4efb\u52a1\u4e5f\u53ef\u4ee5\u53c2\u8003\u8fd9\u4e2a\u601d\u8def\u5b9e\u73b0\u3002\u4f8b\u5982\u56fe\u50cf\u8bed\u4e49\u5206\u5272\u4e2d\u8ba1\u7b97mIoU\u53ea\u9700\u8981\u5c06\u6bcf\u4e2a\u8fdb\u7a0b\u7684\u6df7\u6dc6\u77e9\u9635\u6c47\u603b\u76f8\u52a0\u5230rank0\u5373\u53ef\u3002<\/p>\n\n\n\n<h2>\u4e00\u4e9b\u53ef\u80fd\u9047\u5230\u7684\u95ee\u9898<\/h2>\n\n\n\n<p><a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E4%B8%80%E4%BA%9B%E5%8F%AF%E8%83%BD%E9%81%87%E5%88%B0%E7%9A%84%E9%97%AE%E9%A2%98\"><\/a><\/p>\n\n\n\n<p>\u7f51\u7edc\u9632\u706b\u5899\u6709\u53ef\u80fd\u5728\u9996\u6b21\u591a\u673a\u591a\u5361\u8bad\u7ec3\u65f6\u9020\u6210\u8ba1\u7b97\u8282\u70b9\u95f4\u7684\u901a\u4fe1\u5931\u8d25\u3002\u5355\u673a\u591a\u5361\u6210\u529f\u8fd0\u884c\u7684\u4ee3\u7801\u5728\u6269\u5c55\u81f3\u591a\u673a\u591a\u5361\u9047\u5230\u95ee\u9898\u540e\u53ef\u4ee5\u9996\u5148\u5c1d\u8bd5\u5c06init_method\u5207\u6362\u4e3aGloo\uff0c\u80fd\u591f\u56de\u907f\u6389\u4e00\u4e9b\u6f5c\u5728\u7684\u95ee\u9898\u3002\u8bb0\u5f55\u4e00\u4e0b\u672c\u4eba\u5728\u5b9e\u8df5\u4e2d\u9047\u5230\u7684\u95ee\u9898\u548c\u89e3\u51b3\u65b9\u6cd5\u3002<\/p>\n\n\n\n<h3><a href=\"https:\/\/discuss.pytorch.org\/t\/runtimeerror-address-family-mismatch-when-use-gloo-backend\/64753\">address family mismatch \u9519\u8bef<\/a><\/h3>\n\n\n\n<p><a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#address-family-mismatch-%E9%94%99%E8%AF%AF\"><\/a><\/p>\n\n\n\n<p>\u89e3\u51b3\u65b9\u6848\u662f\u624b\u52a8\u8bbe\u7f6e\u901a\u4fe1\u7684\u7f51\u7edc\u7aef\u53e3\u3002\u673a\u5668\u7684\u7f51\u7edc\u7aef\u53e3\u901a\u8fc7<code>ifconfig<\/code>\u547d\u4ee4\u67e5\u8be2\uff0c\u6709\u591a\u4e2a\u7f51\u53e3\u65f6\u53ef\u4ee5\u90fd\u5c1d\u8bd5\u4e00\u4e0b\u3002<\/p>\n\n\n\n<p>\u5f53backend==NCCL<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\"># Node 0 \n# terminal-0\nexport NCCL_SOCKET_IFNAME=eth0\npython ...\n\n# Node 1 : \n# terminal-0\nexport NCCL_SOCKET_IFNAME=eth0\npython ...<\/pre>\n\n\n\n<p>\u5f53backend==Gloo<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\"># Node 0 \n# terminal-0\nexport GLOO_SOCKET_IFNAME=eth0\npython ...\n\n# Node 1 : \n# terminal-0\nexport GLOO_SOCKET_IFNAME=eth0\npython ...<\/pre>\n\n\n\n<h2>\u53c2\u8003<\/h2>\n\n\n\n<p><a href=\"https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP-Tutorial.md#%E5%8F%82%E8%80%83\"><\/a><\/p>\n\n\n\n<ol><li><a href=\"https:\/\/pytorch.org\/docs\/stable\/distributed.html#choosing-the-network-interface-to-use\">https:\/\/pytorch.org\/docs\/stable\/distributed.html#choosing-the-network-interface-to-use<\/a><\/li><li><a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/dist_overview.html\">https:\/\/pytorch.org\/tutorials\/beginner\/dist_overview.html<\/a><\/li><li>Li, S., Zhao, Y., Varma, R., Salpekar, O., Noordhuis, P., Li, T., &#8230; &amp; Chintala, S. (2020). Pytorch distributed: Experiences on accelerating data parallel training. arXiv preprint arXiv:2006.15704.<\/li><li><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/76638962\">https:\/\/zhuanlan.zhihu.com\/p\/76638962<\/a><\/li><li><a href=\"https:\/\/yangkky.github.io\/2019\/07\/08\/distributed-pytorch-tutorial.html\">https:\/\/yangkky.github.io\/2019\/07\/08\/distributed-pytorch-tutorial.html<\/a><\/li><li><a href=\"https:\/\/medium.com\/huggingface\/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255\">https:\/\/medium.com\/huggingface\/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255<\/a><\/li><\/ol>\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/github.com\/KaiiZhang\/DDP-Tutorial\/blob\/main\/DDP &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2024\/07\/17\/pytorchdistributeddataparallel\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">pytorch\u591a\u673a\u591a\u5361\u8bad\u7ec3\u3010DistributedDataParallel\u3011<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11,4,39],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/16283"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=16283"}],"version-history":[{"count":24,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/16283\/revisions"}],"predecessor-version":[{"id":16311,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/16283\/revisions\/16311"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=16283"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=16283"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=16283"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}