{"id":19021,"date":"2024-09-18T19:49:20","date_gmt":"2024-09-18T11:49:20","guid":{"rendered":"http:\/\/139.9.1.231\/?p=19021"},"modified":"2024-12-31T10:50:38","modified_gmt":"2024-12-31T02:50:38","slug":"discuss-how-to-train-a-model-using-1024-gpus","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/09\/18\/discuss-how-to-train-a-model-using-1024-gpus\/","title":{"rendered":"\u8ba8\u8bba\uff1a\u5982\u4f55\u75281024\u5f20\u663e\u5361\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b?"},"content":{"rendered":"\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e00\u4e2a<a rel=\"noreferrer noopener\" href=\"https:\/\/www.zhihu.com\/question\/650979052\" target=\"_blank\"><strong>\u77e5\u4e4e\u95ee\u9898:\u5982\u4f55\u5224\u65ad\u5019\u9009\u4eba\u6709\u6ca1\u6709\u5343\u5361GPU\u96c6\u7fa4\u7684\u8bad\u7ec3\u7ecf\u9a8c\uff1f<\/strong><\/a>\u786e\u5b9e\u5bf9\u4e8e\u666e\u901a\u5f00\u53d1\u8005\u6765\u8bf4\uff0c\u5927\u6982\u7387\u4ece\u672a\u5c1d\u8bd5\u8fc7\u4f7f\u7528\u6570\u5343\u5f20GPU\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\uff0c\u8fd9\u65b9\u9762\u786e\u5b9e\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u7814\u7a76\u65b9\u5411\uff0c\u4e5f\u662f\u6210\u4e3a\u9876\u5c16\u7b97\u6cd5\u5de5\u7a0b\u5e08\u6240\u5fc5\u9700\u7684\u5fc5\u7ecf\u4e4b\u8def\uff0c\u56e0\u6b64\u8bb0\u5f55\u4e0b\u77e5\u4e4e\u7684\u4e00\u4e9b\u56de\u7b54\uff0c\u7528\u4e8e\u5b66\u4e60\u548c\u8bb0\u5f55\u3002\u867d\u7136\u76ee\u524d\u8fd8\u6ca1\u6709\u673a\u4f1a\u80fd\u591f\u8c03\u7528\u6570\u5343\u5f20GPU\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\uff0c\u4f46\u5bf9\u4e8e\u76ee\u524d\u51e0\u5341\u5f20GPU\u8fdb\u884c\u5e76\u884c\u8bad\u7ec3\u4e5f\u6709\u5e2e\u52a9\u3002<\/p>\n\n\n\n\n\n<h2>\u9ad8\u8d5e\u56de\u7b541\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/zyc.ai\/sketch\/document\/train_nn_with_1024gpus\/\" target=\"_blank\">\u5982\u4f55\u7528\u5343\u5361\u8fdb\u884c\u8bad\u7ec3<\/a><\/h2>\n\n\n\n<p>\u6700\u8fd1\u770b\u5230\u77e5\u4e4e\u4e00\u4e2a<a href=\"https:\/\/www.zhihu.com\/question\/650979052\/answer\/3454365035\">\u56de\u7b54<\/a>\uff0c\u628a\u5343\u5361\u8bad\u7ec3\u7684\u96be\u5ea6\u5439\u4e0a\u5929\u4e86<strong>\u3002\u4f46\u5176\u5b9e\u771f\u6b63\u7528\u8fc7\u5343\u5361\u5c31\u4f1a\u53d1\u73b0\u4e5f\u5c31\u90a3\u4e48\u51e0\u4e2a\u70b9<\/strong>\u3002\u4e8e\u662f\u60f3\u5199\u4e00\u7bc7\u6587\u7ae0\u7b80\u5355\u8bb2\u8bb2\u3002<\/p>\n\n\n\n<p class=\"has-bright-blue-background-color has-background\"><strong>\u672c\u6587\u5c06\u5305\u62ec\u4e09\u4e2a\u90e8\u5206\uff1a<\/strong><\/p>\n\n\n\n<ul><li><strong>\u9996\u5148\u6211\u4eec\u5c06\u8ba8\u8bba\u5343\u5361\u8bad\u7ec3\u7684\u96be\u9898\uff0c\u4ee5\u53ca\u5e94\u8be5\u5728\u4ec0\u4e48\u65f6\u5019\u4f7f\u7528\u5343\u5361\u8bad\u7ec3\uff1b<\/strong><\/li><li><strong>\u63a5\u7740\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u5982\u4f55\u5728\u4e00\u5343\u5f20\u5361\u4e0a\u5f00\u59cb\u8bad\u7ec3\uff0c\u5982\u4f55\u8ba9\u4ed6\u8fbe\u5230\u8fd1\u4e4e\u7ebf\u6027\u7684\u6027\u80fd\u63d0\u5347\uff1b<\/strong><\/li><li><strong>\u6700\u540e\u6211\u4eec\u5c06\u5c55\u5f00\u8ba8\u8bba\u4e00\u4e9b\u5343\u5361\u8bad\u7ec3\u5f53\u4e2d\u4ecd\u7136\u60ac\u800c\u672a\u51b3\uff08\u81f3\u5c11\u5bf9\u4e8e\u5f00\u6e90\u793e\u533a\u6765\u8bf4\uff09\u7684\u95ee\u9898\u3002<\/strong><\/li><\/ul>\n\n\n\n<h3 id=\"_1\">\u4e3a\u4ec0\u4e48\u5343\u5361\u8bad\u7ec3\u662f\u56f0\u96be\u7684\uff1f<\/h3>\n\n\n\n<p>\u5176\u5b9e\u90a3\u7bc7\u56de\u7b54\u5728\u8fd9\u90e8\u5206\u8bf4\u7684\u6ca1\u9519\u3002\u5343\u5361\u8bad\u7ec3\u548c\u516b\u5361\u8bad\u7ec3\u7684\u533a\u522b\u662f&#8212;\u663e\u5361\u591a\u4e86\u4e00\u767e\u591a\u500d\u3002<\/p>\n\n\n\n<p>\u8fd9\u610f\u5473\u7740\u4ec0\u4e48\u5462\uff1f<\/p>\n\n\n\n<ol><li>\u901a\u4fe1\u65f6\u95f4\u589e\u52a0<\/li><li>\u6545\u969c\u6982\u7387\u589e\u52a0<\/li><\/ol>\n\n\n\n<p>\u8fd9\u4fe9\u95ee\u9898\u90fd\u5f88\u597d\u7406\u89e3\u3002<\/p>\n\n\n\n<p>\u65f6\u95f4\u4e0a\uff0cPyTorch \u5185\u90e8\u652f\u6301 NCCL \/ Gloo \/ MPI \u4e09\u4e2a\u901a\u4fe1\u540e\u7aef\uff08<strong>\u8bf7\u52a1\u5fc5\u4f7f\u7528 NCCL\u3002<\/strong>)\u5176\u4e2d\u8bad\u7f51\u7edc\u6700\u5e38\u7528\u7684 AllReduce \u64cd\u4f5c<sub><strong><em>\u3010\u4ece\u591a\u4e2asender\u90a3\u91cc\u63a5\u6536\u6570\u636e\uff0c\u6700\u7ec8combine\u5230\u4e00\u4e2a\u8282\u70b9\u4e0a\u3011<\/em><\/strong><\/sub>\u4f1a\u6839\u636e\u5177\u4f53\u786c\u4ef6\u914d\u7f6e\u8d70 Ring AllReduce \u548c Tree AllReduce<em><sub>\u3010<a href=\"https:\/\/www.zhihu.com\/question\/57799212\">ring allreduce\u548ctree allreduce\u7684\u5177\u4f53\u533a\u522b\u662f\u4ec0\u4e48\uff1f<\/a>\u3011<\/sub><\/em>\u3002Ring \u7684\u65f6\u95f4\u590d\u6742\u5ea6\u662f&nbsp;<em>O(pn)<\/em>\uff0cTree \u7684\u65f6\u95f4\u590d\u6742\u5ea6\u662f<em>&nbsp;O(log\u2061pn)<\/em>\u3002\u5c31\u7b97\u662f\u7406\u8bba\u4e0a 128 \u8282\u70b9\u4e5f\u6bd4\u5355\u8282\u70b9\u6162\u81f3\u5c11\u4e03\u500d\uff0c\u5b9e\u8df5\u5f53\u4e2d\u8de8\u8282\u70b9\u901a\u4fe1\u8981\u8fdc\u6bd4\u5355\u8282\u70b9\u6162\u5f97\u591a\u3002<\/p>\n\n\n\n<p><strong>\u6545\u969c<\/strong>\u4e0a\uff0c\u4e00\u4e2a\u8282\u70b9\u51fa\u95ee\u9898\u7684\u6982\u7387\u662f<em>&nbsp;p<\/em>\uff0c128 \u4e2a\u8282\u70b9\u5c31\u662f&nbsp;1\u2212(1\u2212<em>p<\/em><sup>128<\/sup>)\u3002\u4e5f\u5c31\u662f\u8bf4\u5982\u679c\u4e00\u4e2a\u64cd\u4f5c\u5728\u4e00\u4e2a\u8bad\u7ec3\u5f53\u4e2d\u7684\u51fa\u9519\u6982\u7387\u662f 1%\uff0c\u90a3\u4e48\u5728 128 \u8282\u70b9\u5f53\u4e2d\u7684\u51fa\u9519\u6982\u7387\u5c31\u662f 72.37%\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u968f\u7740\u89c4\u6a21\u7684\u589e\u5927\uff0c\u8bb8\u591a\u95ee\u9898\u90fd\u4f1a\u53d8\u5f97\u96be\u4ee5\u5fcd\u53d7\u3002<em><strong>\u6bd4\u5982\u6570\u636e\u589e\u5f3a\u8981\u82b1 0.1s\uff0c\u4e00\u4ebf\u6761\u6570\u636e\u5c31\u662f 278 \u4e2a\u5c0f\u65f6\uff08\u5f53\u7136\u8fd9\u53ea\u662f\u80e1\u62c6\u7684\u4e00\u4e2a\u6570\u5b57\uff0c\u5b9e\u9645\u6709\u5404\u79cd\u673a\u5236\u6240\u4ee5\u4e0d\u4f1a\u6709\u8fd9\u4e48\u5927\u5f71\u54cd\u3002<\/strong><\/em><\/p>\n\n\n\n<p><strong>\u56e0\u6b64\uff0c\u94b1\u591a\u70e7\u624b\u5e76\u4e0d\u662f\u4f7f\u7528\u5343\u5361\u8bad\u7ec3\u7684\u7406\u7531\u3002\u95f2\u5f97\u86cb\u75bc\u53ef\u80fd\u662f\uff0c\u4f46\u4f60\u5f97\u591a\u86cb\u75bc\u624d\u80fd\u60f3\u51fa\u8fd9\u4e48\u6298\u78e8\u81ea\u5df1\u7684 idea\uff1f<\/strong><\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c<strong>\u5343\u5361\u8bad\u7ec3\u89e3\u51b3\u7684\u95ee\u9898\u662f\u5927\u6a21\u578b&amp;\u5927\u6570\u636e\u95ee\u9898<\/strong>\u3002<strong>\u5982\u679c\u4f60\u7684\u8bad\u7ec3\u65f6\u95f4\u6ca1\u6709\u8d85\u8fc7 8192 GPU \u65e5\uff0c\u90a3\u4e48\u4f60\u7edd\u5bf9\u4e0d\u9700\u8981\u4e00\u5343\u5f20\u663e\u5361\u3002<\/strong><\/p>\n\n\n\n<p>\u770b\u5230\u8fd9\u91cc\uff0c\u7edd\u5927\u591a\u6570\u4eba\u5df2\u7ecf\u53ef\u4ee5\u5173\u6389\u8fd9\u7bc7\u6587\u7ae0\u4e86\u3002\u9664\u975e\u4f60\u7684\u6a21\u578b\u548c\u6570\u636e\u90fd\u4ee5 B\uff08\u5341\u4ebf\uff09\u6765\u4f5c\u4e3a\u8ba1\u91cf\u5355\u4f4d\u3002\u5f53\u7136\u5982\u679c\u4f60\u6b63\u5728\u5395\u6240\u91cc\u624b\u673a\u6ca1\u7535\u60f3\u770b\u70b9\u513f\u4e1c\u897f\u89e3\u95f7\u513f\u7684\u8bdd\uff08\u867d\u7136\u6211\u5f88\u6000\u7591\u662f\u5426\u4f1a\u6709\u4eba\u628a\u4ed6\u6253\u51fa\u6765\u2026\u2026\u90a3\u4e48\u53ef\u4ee5\u7ee7\u7eed\u5f80\u4e0b\u770b<\/p>\n\n\n\n<h3 id=\"_2\">\u5982\u4f55\u4f7f\u7528\u4e00\u5343\u5f20\u5361\u8bad\u7ec3\uff1f<\/h3>\n\n\n\n<h4 id=\"_3\">\u5982\u4f55\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\uff1f<\/h4>\n\n\n\n<p>\u8fd9\u4ef6\u4e8b\u60c5\u5176\u5b9e\u662f\u4e00\u4e2a case by case \u7684\u4e8b\u60c5\u3002\u56e0\u4e3a\u901a\u4fe1\u3001\u8ba1\u7b97\u901f\u5ea6\u5565\u7684\u53d7\u786c\u4ef6\u5f71\u54cd\u66f4\u591a\u3002\u540c\u6837\u662f A100 \u96c6\u7fa4\uff0c\u6211\u5168 DGX \u8282\u70b9\uff0c\u6bcf\u4e00\u5f20 A100 \u90fd\u662f SXM \u63a5\u53e3\u5e76\u914d\u4e00\u5757\u513f\u4e13\u5c5e\u7684 IB \u7f51\u5361\u3002\u4f60\u4e00\u4e2a\u5c0f\u7834\u666e\u60e0\u670d\u52a1\u5668\u63d2 8 \u5f20 PCI-E A100\uff0cIB \u5361\u4e00\u4e2a\u8282\u70b9\u53ea\u7ed9\u4e00\u5f20\u3002\u90a3\u54b1\u4fe9\u9047\u5230\u7684\u95ee\u9898\u5c31\u5b8c\u5168\u4e0d\u662f\u4e00\u4e2a\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u8981\u8ba8\u8bba\u5982\u4f55\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3001\u51cf\u5c11\u8bad\u7ec3\u8017\u65f6\uff0c\u6211\u4eec\u9996\u5148\u8981\u4e86\u89e3\u8bad\u7ec3\u8017\u65f6\u5728\u54ea\u91cc\u3002\u90a3\u4e48\uff0c\u4e00\u4e2a\u8bad\u7ec3\u6b65\u7684\u8017\u65f6\u5728\u54ea\u91cc\u5462\uff1f\u9700\u8981\u8c28\u8bb0\uff0c\u6ca1\u6709 profile \u7684\u4f18\u5316\u662f\u6ca1\u6709\u610f\u4e49\u7684\u3002<\/p>\n\n\n\n<p>\u4f60\u53ef\u80fd\u4f1a\u8bf4\uff0cforward backward sync\u3002\u5f88\u597d\uff0c\u8fd9\u8bf4\u660e\u4f60\u4e86\u89e3 PyTorch \u7684\u57fa\u672c\u6d41\u7a0b\u3002\u4e0d\u8fc7\u73b0\u5b9e\u5f53\u4e2d\u8981\u590d\u6742\u5f97\u591a\u3002<\/p>\n\n\n\n<ol><li>dataset \u8bfb\u53d6\u6570\u636e\uff0c\u6784\u5efa\u8f93\u51fa<\/li><li>dataloader collate \u6570\u636e\uff0c\u8fdb\u884c\u6570\u636e\u9884\u5904\u7406<\/li><li>\u6a21\u578b forward \u8ba1\u7b97\u8f93\u51fa<\/li><li><strong>loss compute<\/strong><\/li><li><strong>\u6a21\u578b backward \u8ba1\u7b97\u68af\u5ea6<\/strong><\/li><li><strong>\u6a21\u578b sync \u68af\u5ea6<\/strong><\/li><li><strong>\u4f18\u5316\u5668 step \u66f4\u65b0\u6743\u91cd<\/strong><\/li><li><strong>\u6253\u5370 log<\/strong><\/li><\/ol>\n\n\n\n<p>\u5f53\u7136\u8fd9\u662f\u53ef\u4ee5\u65e0\u9650\u7ec6\u5206\u4e0b\u53bb\u7684\uff0c\u4f46\u4e00\u822c\u8fd9\u4e9b\u5c31\u591f\u4e86\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c<strong>\u9664\u4e86 4-7 \u7684\u8017\u65f6\u662f\u771f\u8017\u65f6\uff0c\u5176\u4ed6\u90fd\u9700\u8981\u901a\u8fc7\u5f02\u6b65\u64cd\u4f5c\u6765\u76d6\u6389\u3002\u8fd9\u4e5f\u662f\u6211\u4eec\u7684\u4f18\u5316\u76ee\u6807\u3002<\/strong><\/p>\n\n\n\n<p>\u5f02\u6b65\u6267\u884c\u5728 PyTorch \u7684 dataloader\u3001CUDA \u548c\u5206\u5e03\u5f0f\u5f53\u4e2d\u90fd\u5b58\u5728\u3002\u524d\u8005\u53ef\u4ee5\u901a\u8fc7\u8bbe\u7f6e num_workers \u548c prefetch_count \u4e3a 0 \u6765\u5173\u95ed\uff0c\u540e\u4e24\u8005\u53ef\u4ee5\u901a\u8fc7 cuda.synchornize \u548c dist.barrier \u6765\u6267\u884c\u624b\u52a8\u540c\u6b65\u3002\u5728 profile \u65f6\uff0c\u6211\u4eec\u9700\u8981\u9996\u5148\u9700\u8981\u6d4b\u6574\u4e2a step \u7684\u65f6\u957f\u3002\u7136\u540e\u518d\u5728\u6bcf\u6b21\u6d4b\u91cf\u524d\u6267\u884c\u624b\u52a8\u540c\u6b65\u6765\u8ba1\u7b97\u6bcf\u4e2a\u90e8\u5206\u7684\u65f6\u957f\u3002\u5982\u679c\u524d\u8005\u7684\u603b\u8017\u65f6\u7b49\u4e8e\u540e\u8005 4-7 \u7684\u8017\u65f6\u4e4b\u548c\uff0c\u90a3\u4e48\u901a\u5e38\u4e0d\u9700\u8981\u6267\u884c\u4efb\u4f55\u64cd\u4f5c\u3002\u4f46\u8fd9\u79cd\u60c5\u51b5\u5728\u5343\u5361\u64cd\u4f5c\u4e2d\u51e0\u4e4e\u4e0d\u53ef\u80fd\u53d1\u751f\u3002<\/p>\n\n\n\n<p>\u7b2c 6 \u6b65\u901a\u4fe1\u5f80\u5f80\u9700\u8981\u8017\u8d39\u5927\u91cf\u65f6\u95f4\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8fd8\u9700\u8981\u8fdb\u4e00\u6b65\u4f18\u5316\u901a\u4fe1\u3002<\/p>\n\n\n\n<p>\u4ee5\u4e0b\u5185\u5bb9\u662f\u5bf9<a href=\"https:\/\/arxiv.org\/abs\/2006.15704\"><strong><em>PyTorch Distributed<\/em><\/strong><\/a>\u7684\u6982\u62ec\uff0c\u6709\u611f\u5174\u8da3\u7684\u540c\u5b66\u5efa\u8bae\u901a\u8bfb\u5e76\u80cc\u8bf5\u5168\u6587\u3002<\/p>\n\n\n\n<h5 id=\"-\">\u8ba1\u7b97-\u901a\u4fe1\u91cd\u53e0<\/h5>\n\n\n\n<p>\u5728 PyTorch \u5f53\u4e2d\uff0c\u68af\u5ea6\u7684\u901a\u4fe1\u548c\u53cd\u5411\u4f20\u64ad\u662f\u4ea4\u53e0\u8fdb\u884c\u7684\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c<strong>\u6bcf\u5b8c\u6210\u4e00\u5c42\u7684\u68af\u5ea6\u8ba1\u7b97\uff0c\u90fd\u4f1a\u7acb\u5373\u89e6\u53d1\u5f53\u524d\u5c42\u7684\u540c\u6b65\u3002\u5b9e\u73b0\u8d77\u6765\u4e5f\u5f88\u7b80\u5355\uff0c\u6bcf\u4e2a\u8fdb\u7a0b\u5728\u5b8c\u6210\u81ea\u5df1\u7b2c k \u5c42\u7684\u68af\u5ea6\u8ba1\u7b97\u540e\u90fd\u4f1a\u89e6\u53d1\u4e00\u4e2a\u94a9\u5b50\u6765\u7ed9\u8ba1\u6570\u5668+1s\u3002\u5f53\u8ba1\u6570\u5668\u8fbe\u5230\u8fdb\u7a0b\u6570\u65f6\u5f00\u706b\u8fdb\u884c\u68af\u5ea6\u901a\u4fe1\u3002<\/strong>\u6709\u5f88\u591a\u540c\u5b66\u5728\u8ba1\u7b97\u68af\u5ea6\u8fc7\u7a0b\u4e2d\u9047\u5230\u8fc7\u00a0<code>RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one.<\/code>\u00a0\u9519\u8bef\uff0c\u8fd9\u5c31\u662f\u56e0\u4e3a\u6709\u7684\u6a21\u5757\u6ca1\u6709\u53c2\u4e0e\u8ba1\u7b97 loss\uff0c\u5bfc\u81f4\u68af\u5ea6\u540c\u6b65\u5361\u4f4f\u4e86\u3002\u9700\u8981\u6ce8\u610f\uff0c\u5f53\u00a0<code>find_unused_parameters=True<\/code>\u00a0\u65f6\uff0cPyTorch \u5206\u5e03\u5f0f\u4f7f\u7528\u00a0<code>nn.Module.__init__<\/code>\u00a0\u5f53\u4e2d\u5b9a\u4e49\u5b50\u6a21\u5757\u7684\u53cd\u5411\u987a\u5e8f\u6765\u4f5c\u4e3a\u68af\u5ea6\u6876\u7684\u6784\u5efa\u987a\u5e8f\u3002\u56e0\u6b64\uff0c\u786e\u4fdd\u6a21\u5757\u5b9a\u4e49\u548c\u8c03\u7528\u7684\u987a\u5e8f\u4e00\u81f4\u662f\u4e00\u4e2a\u826f\u597d\u7684\u5b9e\u8df5\u3002<\/p>\n\n\n\n<h5 id=\"_4\">\u68af\u5ea6\u5408\u6876<\/h5>\n\n\n\n<p>\u5c3d\u7ba1\u7406\u8bba\u4e0a\u6765\u8bf4\uff0c<strong>\u540c\u6b65\u53d1\u751f\u7684\u8d8a\u53ca\u65f6\uff0c\u91cd\u5408\u5ea6\u8d8a\u9ad8\uff0c\u6027\u80fd\u8d8a\u597d\u3002<\/strong>\u4f46\u5b9e\u9645\u4e0a\u6bcf\u6b21\u53d1\u8d77\u901a\u4fe1\u90fd\u662f\u6709\u4e0a\u5934\u7684\u3002\u56e0\u6b64\uff0c\u73b0\u5b9e\u5f53\u4e2d\u68af\u5ea6\u540c\u6b65\u5e76\u4e0d\u662f\u8d8a\u591a\u8d8a\u597d\u8d8a\u5feb\u8d8a\u597d\u3002\u4e3a\u6b64\uff0cPyTorch \u5f15\u5165\u4e86<strong>\u68af\u5ea6\u5408\u6876\u673a\u5236<\/strong>\uff0c\u901a\u8fc7\u628a\u591a\u4e2a Tensor \u88c5\u5728\u4e00\u4e2a\u6876\u91cc\u518d\u901a\u4fe1\u6876\u6765\u51cf\u5c11\u901a\u4fe1\u6b21\u6570\u4ece\u800c\u51cf\u5c11\u603b\u8017\u65f6\u3002<strong>\u5408\u6876\u7684\u00a0<code>bucket_cap_mb<\/code>\u00a0\u9ed8\u8ba4\u662f 25MiB<\/strong>\uff0c\u8fd9\u5bf9\u4e8e\u7edd\u5927\u591a\u6570\u6a21\u578b\u6765\u8bf4\u90fd\u662f\u592a\u5c0f\u7684\u3002\u76ee\u524d\u5df2\u7ecf\u6709\u63d0\u5347\u8fd9\u4e2a\u9ed8\u8ba4\u503c\u7684<a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/118421\">\u7279\u6027\u9700\u6c42<\/a>\uff0c\u4f46\u662f\u8fd9\u4e2a\u8fd8\u662f\u8c03\u4e00\u4e0b\u66f4\u597d\u3002<\/p>\n\n\n\n<h5 id=\"_5\">\u68af\u5ea6\u7d2f\u52a0<\/h5>\n\n\n\n<p>\u5f53\u4f60\u505a\u5b8c\u6240\u6709\u64cd\u4f5c\u4e4b\u540e\uff0c\u60ca\u559c\u7684\u53d1\u73b0 TMD \u600e\u4e48\u540c\u6b65\u65f6\u95f4\u8fd8\u662f\u5355\u8282\u70b9\u7684\u597d\u51e0\u500d\u3002\u8fd9\u5176\u5b9e\u662f\u6b63\u5e38\u60c5\u51b5\u2026\u2026<strong>\u5b9e\u9645\u4e0a\u8d85\u8fc7 256 \u5361\u7684\u8bad\u7ec3\u60f3\u8981\u628a\u901a\u4fe1\u76d6\u6389\u5c31\u662f\u4e00\u4ef6\u4e0d\u53ef\u80fd\u7684\u4e8b\u60c5<\/strong>\u3002\u4f60\u8bf4\u8001\u5e08\u6211\u770b FB \u8bba\u6587\u8bf4\u4ed6\u4eec 256 \u5361\u5c31\u662f\u7ebf\u6027\u63d0\u5347\u554a\u2026\u90a3\u8fd9\u91cc\u4e0d\u5f97\u4e0d\u63d0\u7684\u4e00\u4e2a\u7b56\u7565\u5c31\u662f<strong>\u68af\u5ea6\u7d2f\u52a0<\/strong>\u4e86\u3002\u68af\u5ea6\u7d2f\u52a0\u4f1a\u6267\u884c k \u6b21 forward+backward \u4e4b\u540e\u518d\u6267\u884c\u4f18\u5316\u5668\u6b65\u8fdb\u3002\u8fd9\u6709\u5f88\u591a\u597d\u5904\uff0c\u9996\u5148\u5bf9\u4e8e\u5927\u6a21\u578b batch size \u901a\u5e38\u4e0d\u80fd\u5f00\u591a\u5927\uff0c\u68af\u5ea6\u7d2f\u52a0\u53ef\u4ee5\u63d0\u5347\u7b49\u6548 batch size\u3002\u5176\u6b21\u7d2f\u52a0\u671f\u95f4\u7684 backward \u4e0d\u9700\u8981\u901a\u4fe1\u68af\u5ea6\uff0c\u52a0\u5feb\u4e86\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n\n\n\n<h5 id=\"_6\">\u5c11\u5373\u662f\u5feb<\/h5>\n\n\n\n<p>Python \u662f\u4e00\u79cd\u5f88\u6162\u7684\u8bed\u8a00\u3002\u5f53\u7136\u4f60\u8bf4 JIT trace+torch.compile \u6709\u63d0\u5347\u6211\u4e5f\u4e0d\u53cd\u5bf9\uff0c\u4f46<strong>\u5bf9\u4e8e\u6700\u9ad8\u6548\u7387\u6765\u8bf4\uff0c\u53ea\u6709\u5fc5\u987b\u8981\u5b58\u5728\u7684\u4ee3\u7801\u548c\u4e0d\u5b58\u5728\u7684\u4ee3\u7801\u4e24\u79cd\u3002<\/strong><\/p>\n\n\n\n<p>\u62b1\u62b1\u8138\u7684 Transformers \u5c31\u662f\u4e00\u4e2a\u53cd\u4f8b\u3002\u4e24\u4e2a\u5b50\u6a21\u5757\u5c31\u80fd\u5199\u5b8c\u7684 TransformerLayer \u4ed6\u4eec\u786c\u662f\u80fd\u5199\u51fa\u6765\u4e00\u5806\u2026\u504f\u504f\u4ed6\u4eec\u8fd8\u4fe1\u5949 Single Model File Policy\u2026\u2026\u6211\u5bfb\u601d\u4f60\u8fd9\u5b8c\u5168\u4e0d\u8003\u8651\u7ee7\u627f\u7684\u5c01\u8fd9\u4e48\u591a\u5c42\u662f\u8981\u641e\u9e21\u6bdb\u554a\uff1f\u6b63\u4f8b\u53cd\u800c\u662f PyTorch\u2026\u2026\uff08\u7b11\u6b7b\uff0c\u6211\u7adf\u7136\u4f1a\u5938\u8138\u4e66\u4ee3\u7801\u5199\u5f97\u597d\u3002\u5177\u4f53\u6765\u8bf4\u5c31\u662f nn.functional \u5f53\u4e2d\u7684\u5404\u79cd\u5b9e\u73b0\u3002\u4f60\u4f1a\u53d1\u73b0\u4ed6\u4eec\u7b2c\u4e00\u884c\u5f80\u5f80\u662f&nbsp;<code>handle_torch_func<\/code>\u3002\u719f\u6089 Python \u88c5\u9970\u5668\u7684\u5c0f\u4f19\u6c41\u901a\u5e38\u8981\u95ee\u4e86\uff0c\u4e3a\u5565\u8fd9\u91cc\u4e0d\u7528\u4e2a\u88c5\u9970\u5668\u7edf\u4e00\u4e00\u4e0b\uff1f\u56e0\u4e3a\u88c5\u9970\u5668\u4f1a\u5f15\u5165\u989d\u5916\u7684\u51fd\u6570\u8c03\u7528\uff0c\u989d\u5916\u7684\u51fd\u6570\u8c03\u7528\u5c31\u662f\u989d\u5916\u7684\u4e0a\u5934\u3002<\/p>\n\n\n\n<p><strong>\u56e0\u6b64\uff0c\u5982\u679c\u4f60\u60f3\u786e\u4fdd\u6700\u9ad8\u7684\u6548\u7387\uff0c\u5199\u4e00\u4e2a\u7b80\u5355\u7684\u8bad\u7ec3\u4ee3\u7801\u548c\u6a21\u578b\u4ee3\u7801\u975e\u5e38\u91cd\u8981\u3002\u6bd5\u7adf\uff0c1%\u7684\u6548\u7387\u63d0\u5347\uff0c\u8282\u7701\u7684\u53ef\u80fd\u662f\u6570\u767e\u4e2a GPU \u65e5\u3002<\/strong><\/p>\n\n\n\n<h4 id=\"_7\">\u5982\u4f55\u5e73\u7a33\u8bad\u7ec3<\/h4>\n\n\n\n<p><strong>\u8fd9\u4e00\u6bb5\u5f53\u4e2d\u4e2d\u54b1\u4eec\u53ea\u8ba8\u8bba\u4f60\u80fd\u63a7\u5236\u7684\u95ee\u9898\u3002<\/strong><\/p>\n\n\n\n<h5 id=\"_8\">\u6355\u6349\u4e0d\u81f4\u547d\u7684\u5f02\u5e38<\/h5>\n\n\n\n<p><strong>\u6545\u969c\u7387\u9ad8\u7684\u95ee\u9898\u5176\u5b9e\u5f88\u597d\u89e3\u51b3\u3002\u5728\u8bad\u7ec3\u5f53\u4e2d\uff0c\u5927\u90e8\u5206\u5f02\u5e38\u90fd\u662f\u975e\u81f4\u547d\u5f02\u5e38\uff0c\u63a5\u4f4f\u4ed6\u4eec\u5c31\u597d\u4e86\u3002\u6211\u4e4b\u524d\u5199\u8fc7\u4e00\u4e2a\u88c5\u9970\u5668\uff0c<a href=\"https:\/\/danling.org\/utils\/decorators\/#danling.utils.decorators.catch\">catch<\/a>\uff0c\u5b83\u7684\u4f5c\u7528\u5c31\u662f\u63a5\u4f4f\u5f02\u5e38\uff0c\u7136\u540e\u8c03\u56de\u8c03\u51fd\u6570\uff08\u9ed8\u8ba4\u5f53\u7136\u5c31\u662f\u628a\u9519\u8bef\u6253\u5370\u5230 log \u91cc\uff09\u3002\u6240\u6709\u4f60\u9700\u8981\u505a\u7684\u53ea\u662f\u4f7f\u7528\u5b83\u6765\u88c5\u9970\u6240\u6709\u975e fatal \u7684\u64cd\u4f5c\u3002<\/strong><\/p>\n\n\n\n<p><strong>\u5728\u5b9e\u9645\u5e94\u7528\u5f53\u4e2d\uff0c\u6211\u4eec\u9047\u5230\u7684\u6700\u5e38\u89c1\u7684\u95ee\u9898\u662f\u5b58 ckpt \u5199\u6ee1\u4e86\u78c1\u76d8\uff08\u4e0d\u51c6\u7b11\uff0c\u4ece\u5546\u6c64\u5230\u4e0a\u6d77 AI Lab\uff0c\u8fd9\u4e2a\u95ee\u9898\u5728\u54ea\u513f\u90fd\u65e5\u5e38\u51fa\u73b0\u3002\u54b1\u4e5f\u4e0d\u77e5\u9053\u4e3a\u5565\u80af\u4e70\u90a3\u4e48\u591a\u663e\u5361\u4f46\u4e0d\u80af\u591a\u63d2\u70b9\u513f\u786c\u76d8\uff0c\u54b1\u4e5f\u4e0d\u6562\u95ee\uff09\u3002<\/strong>\u63a5\u4f4f\u6240\u6709\u4fdd\u5b58\u64cd\u4f5c\uff0c\u5982\u679c\u4f60\u6709\u95f2\u5fc3\u53ef\u4ee5\u5728\u56de\u8c03\u91cc\u5220\u4e00\u4e0b\u4e4b\u524d\u7684 ckpt\u3002\u6ca1\u95f2\u5fc3\u7684\u8bdd\u2026\u5927\u4e0d\u4e86\u91cd\u8bad\u4e00\u6b21\u561b\uff08\u9003\u3002\uff09\u7b2c\u4e8c\u5e38\u89c1\u7684\u95ee\u9898\uff0c\u4f60\u731c\u5bf9\u4e86\u2026\u2026\u5b58 log \u5199\u6ee1\u4e86\u786c\u76d8\u2026\u2026\u6240\u4ee5\u6240\u6709 logging \u64cd\u4f5c\u4e5f\u90fd\u662f\u8981 catch \u7684\u3002<strong>\u8fd9\u5c31\u662f\u4e3a\u5565\u6211\u90fd\u7528 tmux \u7136\u540e\u5f00\u5f88\u957f\u7684\u7f13\u5b58\u7a97\u53e3\uff0c\u603b\u662f\u80fd\u62a2\u6551\u4e00\u4e9b log \u51fa\u6765\u7684\u3002<\/strong><\/p>\n\n\n\n<p>\u54b3\u54b3\uff0c\u8bf4\u70b9\u513f\u6b63\u7ecf\u7684\u3002\u4efb\u4f55\u8054\u7f51\u64cd\u4f5c\u90fd\u662f\u9700\u8981 catch \u7684\uff0c\u5e38\u89c1\u7684\u8054\u7f51\u64cd\u4f5c\u4e3b\u8981\u5305\u62ec\u4ece ceph \u8bfb\u53d6\u6570\u636e\u548c\u2026\u5199 log \u5230\u8fdc\u7a0b\uff08\u9003\u3002\u5176\u4ed6\u5c31\u6ca1\u5565\u4e86\u5427\uff0c\u6211\u89c1\u8fc7\u6709\u5927\u54e5\u5c1d\u8bd5\u6062\u590d OOM \u7684\uff0c\u4f46\u6548\u679c\u4f3c\u4e4e\u4e0d\u662f\u5f88\u597d\uff0c\u81f3\u5c11\u6211\u81ea\u5df1\u6ca1\u7528\u8fc7\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u552f\u4e00\u4e0d\u5e94\u6355\u6349\u7684\u9519\u8bef\u662f\u96c6\u7fa4\u70b8\u4e86\u3002<\/p>\n\n\n\n<p>\u90a3\u6709\u7684\u5927\u5144\u5f1f\u5c31\u8bf4\u4e86\uff0c\u96c6\u7fa4\u6ca1\u7206\u70b8\uff0c\u4f46\u662f\u6709\u4e24\u5f20\u5361\u7a81\u7136\u6389\u4e86\u548b\u529e\u3002\u8fd9\u4e2a\u54b1\u7b2c\u4e09\u90e8\u5206\u518d\u8ba8\u8bba\u3002<\/p>\n\n\n\n<h5 id=\"_9\">\u7ba1\u597d\u6a21\u578b\u7684\u8f93\u51fa<\/h5>\n\n\n\n<p><strong>\u6a21\u578b\u8bad\u7740\u8bad\u7740\u53d1\u6563\u4e86\u51e0\u4e4e\u662f\u6bcf\u4e2a\u8bad\u5927\u6a21\u578b\u7684\u4eba\u90fd\u4f1a\u9047\u5230\u7684\u95ee\u9898\u3002\u8f93\u51fa\u548c loss \u53ea\u8981\u6709 nan \u679c\u65ad\u4e22\u6389\u3002<\/strong>\u68af\u5ea6\u5148 clip by value \u518d clip by norm \u90fd\u662f\u5e38\u89c4\u64cd\u4f5c\u3002\u54e6\u5bf9\u4e86\uff0c\u8fd8\u6709\u521d\u59cb\u5316\u2026\u2026\u5173\u4e8e\u5927\u6a21\u578b\u6536\u655b\u6027\u7684\u8bba\u6587\u6709\u4e00\u5806\uff0c\u6b64\u5904\u4e0d\u518d\u8d58\u8ff0\u3002<\/p>\n\n\n\n<h4 id=\"_10\">\u6bd4\u66f4\u5927\uff0c\u8fd8\u66f4\u5927\uff0c\u518d\u66f4\u5927<\/h4>\n\n\n\n<h5 id=\"_11\">\u5f39\u6027\u8bad\u7ec3<\/h5>\n\n\n\n<p><strong>\u5b9e\u9645\u4e0a\u5f53\u4f60\u7684\u8bad\u7ec3\u8d85\u8fc7 2048 \u4e2a GPU \u65e5\u65f6\uff0c\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u5f53\u4e2d\u53d1\u751f\u5355\u4e2a GPU \u751a\u81f3\u5355\u4e2a\u8282\u70b9\u4e0b\u7ebf\u662f\u518d\u6b63\u5e38\u4e0d\u8fc7\u7684\u4e8b\u60c5\u4e86\u3002<\/strong><\/p>\n\n\n\n<p>PyTorch \u5728 1.10 \u5c31\u5f15\u5165\u4e86 <strong><a href=\"https:\/\/qiankunli.github.io\/2021\/11\/27\/pytorch_elastic.html\" target=\"_blank\" rel=\"noreferrer noopener\">torchelastic \u5f39\u6027\u8bad\u7ec3\u673a\u5236<\/a><\/strong>\u3002\u8fd9\u4e2a\u4e1c\u897f\uff0c\u7528\u8fc7\u7684\u90fd\u9a82\u5a18\u3002\u7b49\u4e0b\uff0c\u8ba9\u6211\u5148\u9a82\u4e00\u904d\u3002\u5478\u3002ok \u54b1\u4eec\u7ee7\u7eed\u5427\u3002<\/p>\n\n\n\n<p>\u6211\u5370\u8c61\u5f53\u4e2d\u5728\u5fae\u8f6f\u7684\u6700\u540e\u4e00\u8f6e\u9762\u8bd5\u5f53\u4e2d\u88ab\u95ee\u5230\u4e86\u8fd9\u4e2a\u95ee\u9898\uff1a\u5982\u4f55\u8bbe\u8ba1\u4e00\u4e2a\u5f39\u6027\u5206\u5e03\u5f0f\u7cfb\u7edf\u3002<\/p>\n\n\n\n<p>\u6211\u7684\u56de\u7b54\u5f88\u6559\u79d1\u4e66\u3002<strong>\u6bcf k \u5206\u949f\uff0c\u7cfb\u7edf\u4f1a\u505a\u4e00\u6b21 AllGather \u6765\u7edf\u8ba1\u5b58\u6d3b\u8fdb\u7a0b\u6570\uff0c\u7136\u540e\u9009\u4e3e\u51fa\u4e00\u4e2a\u4e3b\u8fdb\u7a0b\u3002\u4e3b\u8fdb\u7a0b\u4f1a\u8ba1\u7b97\u597d\u6bcf\u4e2a\u8fdb\u7a0b\u7684 rank \u548c local rank \u7136\u540e\u5e7f\u64ad\u7ed9\u5404\u4e2a\u8fdb\u7a0b\u3002\u6240\u6709\u8fdb\u7a0b\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u5f00\u59cb\u65f6\u5411\u4e3b\u8fdb\u7a0b\u53d1\u9001\u4e00\u4e2a\u5fc3\u8df3\u5305\u6765\u6c47\u62a5\u72b6\u6001\u3002\u4e3b\u8fdb\u7a0b\u4f1a\u6839\u636e\u5fc3\u8df3\u5305\u6765\u786e\u5b9a\u8fd9\u4e00\u4e2a step \u53c2\u4e0e\u540c\u6b65\u7684\u673a\u5668\u6709\u591a\u5c11\u3002<\/strong><\/p>\n\n\n\n<p>\u4f46\u5f88\u53ef\u60dc\uff0c2024 \u5e74\u4e86\u3002\u8fd8\u662f\u6ca1\u4eba\u53bb\u5199\u3002\u4ed6\u5988\u7684\u3002<\/p>\n\n\n\n<h5 id=\"_12\">\u5c42\u6b21\u5316\u68af\u5ea6\u540c\u6b65<\/h5>\n\n\n\n<p>\u6211\u4e00\u76f4\u8ba4\u4e3a\u68af\u5ea6\u540c\u6b65\u4e0d\u5e94\u8be5\u4ee5 GPU\/\u8fdb\u7a0b\u4e3a\u5355\u4f4d\u3002\u800c\u5e94\u8be5\u5206\u4e3a\u5927\u540c\u6b65\uff08\u8282\u70b9\u95f4\u540c\u6b65\uff09\u548c\u5c0f\u540c\u6b65\uff08\u8282\u70b9\u5185\u540c\u6b65\uff09\u3002\u5c0f\u540c\u6b65\u53ef\u4ee5\u66f4\u9ad8\u9891\u7684\u8fdb\u884c\uff0c\u5927\u540c\u6b65\u5219\u53ef\u4ee5\u66f4\u6162\u7684\u6267\u884c\u3002\u8fd9\u6837\u4e0d\u4ec5\u80fd\u63d0\u9ad8\u5b9e\u9645\u7684\u68af\u5ea6\u540c\u6b65\u9891\u7387\uff0c\u964d\u4f4e\u540c\u6b65\u603b\u8017\u65f6\uff0c\u5e76\u4e14\u8fd8\u80fd\u5929\u7136\u7684\u53bb\u7ed3\u5408\u5c0f batch \u548c\u5927 batch \u8bad\u7ec3\u7684\u4f18\u70b9\u2014\u8282\u70b9\u5185\u5c0f batch \u5173\u6ce8\u4e2a\u4f53\uff0c\u8282\u70b9\u95f4\u5927 batch \u5173\u6ce8\u6574\u4f53\u3002<\/p>\n\n\n\n<p class=\"has-light-gray-background-color has-background\"><strong>\u6709\u6ca1\u6709\u53d1\u73b0\u6240\u6709\u4e1c\u897f\u90fd\u5f88\u7b80\u5355\uff1f<br>\u662f\u8fd9\u6837\u7684\uff0c\u5343\u5361\u8bad\u7ec3\u662f\u4efb\u4f55\u4e00\u4e2a\u666e\u901aCS\u672c\u79d1\u751f\u82b1\u4e09\u4e2a\u6708\u5c31\u80fd\u5b66\u4f1a\u7684\u4e1c\u897f\u3002\u6ca1\u6709\u4efb\u4f55\u590d\u6742\u7684\u5730\u65b9\u3002<\/strong><\/p>\n\n\n\n<h3 id=\"_13\">\u5ef6\u4f38\u9605\u8bfb<\/h3>\n\n\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/ddp.html\">PyTorch DDP \u8bbe\u8ba1\u7b14\u8bb0<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/pytorch.org\/tutorials\/recipes\/recipes\/tuning_guide.html\">PyTorch \u5fae\u8c03\u83dc\u8c31<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/varblog.org\/blog\/2018\/05\/24\/profiling-and-optimizing-machine-learning-model-training-with-pytorch\/\">\u5206\u6790\u4e0e\u4f18\u5316\u4f7f\u7528 PyTorch \u8bad\u7ec3\u673a\u5668\u5b66\u4e60\u6a21\u578b<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/dev-discuss.pytorch.org\/t\/using-nsight-systems-to-profile-gpu-workload\/59\">\u4f7f\u7528 Nsight Systems \u6765\u5206\u6790 GPU \u8d1f\u8f7d<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/docs.nvidia.com\/deeplearning\/frameworks\/dlprof-user-guide\/\">DLProf<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/docs.nvidia.com\/cuda\/profiler-users-guide\/index.html\">NVProf<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/github.com\/NVIDIA\/nccl\/issues\/256\">NCCL AllReduce \u8bbe\u8ba1<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/2312.16903\">Spike No More<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/proceedings.mlr.press\/v9\/glorot10a.html\">Understanding the difficulty of training deep feedforward neural networks<\/a><\/p>\n\n\n\n<h2>\u9ad8\u8d5e\u56de\u7b542\uff1a\u5173\u4e8e\u5343\u5361\u8bad\u7ec3\u7684\u96be\u5ea6\u5206\u6790<\/h2>\n\n\n\n<p>\u5343\u5361\uff0c\u5176\u5b9e\u662f\u4e2a\u76f8\u5bf9\u6a21\u7cca\u7684\u6982\u5ff5\uff0c\u5bf9\u4e8e\u591a\u6570\u4eba\u800c\u8a00\uff0c\u8fd9\u5c31\u8ddf\u4f60\u544a\u8bc9\u4f60\uff0c\u6211\u67091\u5146\u8d44\u4ea7\u548c10\u5146\u8d44\u4ea7\u4e00\u6837\uff0c\u4f60\u53ea\u77e5\u9053\u5f88\u591a\u5f88\u591a\uff0c\u4f46\u662f\u5b8c\u5168\u50bb\u50bb\u5206\u4e0d\u6e05\u695a\uff0c\u8fd9\u5230\u5e95\u662f\u591a\u5c11\uff0c\u80fd\u4e70\u591a\u5c11\u4e1c\u897f\u3002\u5343\u5361\uff0c\u4e5f\u662f\u4e00\u6837\u3002\u6309\u7167\u5e38\u89c1\u7684\u4e00\u673a8\u5361GPU\u7684\u7c7b\u578b\u6765\u770b\uff0c\u7528\u8fc7125\u53f0\u673a\u5668\u8bad\u7ec3\u7684\uff0c\u5c31\u7b97\u5f97\u4e0a\u662f\u5343\u5361\u4e86\u3002\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u4e0a\u8bf4\uff0c\u5176\u5b9e\u95e8\u69db\u4e5f\u6ca1\u6709\u90a3\u4e48\u90a3\u4e48\u7684\u9ad8\u3002\u53ef\u4e8b\u5b9e\u5462\uff1f\u771f\u6b63\u7684\u5927\u6a21\u578b\u8981\u7528\u591a\u5c11\u673a\u5668\u8bad\u7ec3\uff1f\u7b54\u6848\u662f\u8fdc\u8d85\u5343\u5361 \u2014\u2014 \u770b\u770b\u4e0b\u9762\u7684GPT-4\u4fe1\u606f\uff0c\u8fd9\u662f\u5343\u5361\u5417\uff1f\u6bd4\u4e07\u5361\u8fd8\u8981\u591a\uff01\uff01\uff01<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>How much compute was used to train GPT-4?<br>2.15e25 floating<\/p><p>Some key facts about how this enormous model was trained: Used 25,000 Nvidia A100 GPUs simultaneously. Trained continuously for 90\u2013100 days. Total compute required was&nbsp;<strong>2.15e25<\/strong>&nbsp;floating point operations.Sep 27, 2023<\/p><\/blockquote>\n\n\n\n<p>\u53e6\u5916\uff0c\u4ece\u6839\u672c\u4e0a\u8bf4\uff0c\u5343\u5361\u8bad\u7ec3\u662f\u4e00\u4e2a\u590d\u5408\u4f53\u3002\u5f88\u591a\u65f6\u5019\uff0c\u5927\u5bb6\u5c31\u77e5\u9053\u4e00\u4ef6\u4e8b\uff0c\u201c\u725b\u903c\u201d\uff0c\u9664\u4e86\u77e5\u9053\u558a666\u4e4b\u5916\uff0c\u5c31\u5c11\u6709\u4e86\u89e3\u5230\u5e95\u725b\u903c\u5728\u54ea\u91cc\u3002\u4ee5\u81f3\u4e8e\u8bf4\uff0c\u89c9\u5f97\u975e\u5e38\u7684\u9ad8\u5927\u4e0a\u3002\u6211\u53ef\u4ee5\u8fd9\u4e48\u8bf4\u4e00\u53e5\uff0c<strong>\u5343\u5361\u53ca\u5176\u4ee5\u4e0a\u7684\u8bad\u7ec3\u5bf9\u4e8e\u7edd\u5927\u591a\u6570\u4eba\u548c\u4f01\u4e1a\u800c\u8a00\uff0c\u8fd9\u5c31\u662f\u4e2a\u5c60\u9f99\u672f<\/strong>&nbsp;\u2014\u2014 \u9664\u4e86\u67d0\u4e9b\u4e2a\uff0c\u7528\u624b\u6307\u5934\u6570\u7684\u51fa\u6765\u7684\u5730\u65b9\uff0c\u522b\u7684\u5730\u65b9\u5b8c\u5168\u6ca1\u6709\u8fd9\u6837\u7684\u9700\u6c42\uff0c\u4e5f\u6ca1\u6709\u8fd9\u6837\u7684\u8d44\u6e90\u6765\u8fdb\u884c\u5343\u5361\u8bad\u7ec3\u3002\u8fd9\u662f\u4ec0\u4e48\u610f\u601d\uff1f\u610f\u601d\u5c31\u662f\u3002\u5982\u679c\u771f\u6709\u8fd9\u6837\u7ecf\u9a8c\u7684\u4eba\uff0c\u6d41\u4e86\u51fa\u6765\uff0c\u5927\u6982\u7387\u5f88\u96be\u5bf9\u53e3\u7684\u627e\u5de5\u4f5c\uff0c\u56e0\u4e3a\u4ed6\u5728\u5343\u5361\u96c6\u8bad\u4e2d\u7684\u8bad\u7ec3\u7ecf\u9a8c\u548c\u5de5\u7a0b\u5b9e\u8df5\uff0c\u5927\u6982\u7387\u6839\u672c\u522b\u7684\u5730\u65b9\u7528\u4e0d\u4e0a\u3002\u53e6\u4e00\u5c42\u610f\u601d\u662f\uff0c\u5982\u679c\u4f60\u53ea\u662f\u4e2a\u4e00\u822c\u4eba\uff0c\u90a3\u4f60\u60f3\u60f3\u5c31\u5f97\u4e86\uff0c\u5c31\u8ddf\u4f60\u53ef\u4ee5\u610f\u6deb\u4e0b\u67d0\u4e2a\u81ea\u5df1\u559c\u6b22\u7684\u660e\u661f\uff0c\u4f46\u522b\u771f\u53bb\u8ffd\u6c42\uff0c\u771f\u8981\u53bb\u4e86\uff0c\u4f60\u5927\u6982\u662f\u8fde\u699c\u4e00\u5927\u54e5\u7684\u5f85\u9047\u90fd\u4e0d\u4f1a\u6709\uff0c\u6ce8\u5b9a\u4e86\u4eba\u8d22\u4e24\u7a7a\u3002\u6211\u5f53\u7136\u77e5\u9053\u6709\u4eba\u4f1a\u8bf4\uff0c\u90a3\u96be\u9053\u6d41\u51fa\u6765\u7684\u4eba\uff0c\u4e0d\u80fd\u53bb\u90a3\u51e0\u4e2a\u6307\u5934\u6570\u7684\u51fa\u6765\u7684\u5730\u65b9\u5417\uff1f\u53ef\u4ee5\u7684\uff0c\u4f46\u662f\u522b\u7740\u6025\uff0c\u4f60\u4eec\u5f80\u540e\u9762\u770b\u5c31\u77e5\u9053\u4e86 \u2014\u2014 \u53bb\u8fd8\u662f\u80fd\u53bb\u7684\uff0c\u4f46\u662f\u5982\u679c\u4ed6\u4e0d\u662f\u8ddf\u7740\u5927\u4f6c\u4e00\u8d77\u8dd1\uff0c\u5230\u4e86\u65b0\u5730\u65b9\u4ed6\u4eec\u771f\u4e0d\u89c1\u7684\u80fd\u7ee7\u7eed\u5e72\u3002<\/p>\n\n\n\n<p>\u518d\u6765\u8bf4\u8bf4\uff0c<strong>\u5343\u5361\u8bad\u7ec3\u662f\u4e2a\u4ec0\u4e48\u590d\u5408\u4f53 \u2014\u2014 \u81f3\u5c11\u662f\uff0c\u79d1\u5b66\uff0c\u5de5\u7a0b\u548c\u4eba\u60c5\u4e16\u6545<\/strong>\u3002\u5148\u770b\u5927\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u4efb\u52a1\u600e\u4e48\u4ece\u201c\u91cf\u53d8\u5230\u8d28\u53d8\u201d\u7684 \u2014\u2014 \u4efb\u4f55\u4e00\u4e2a\u5c0f\u89c4\u6a21\u8bad\u7ec3\u4e0a\u7684\u95ee\u9898\uff0c\u653e\u5927\u51e0\u767e\u51e0\u5343\u500d\u4e4b\u540e\uff0c\u90fd\u6709\u53ef\u80fd\u6210\u4e3a\u4e0d\u53ef\u5ffd\u89c6\u7684\u95ee\u9898\u3002\u6bd4\u5982\uff0c\u6570\u636e\u9884\u5904\u7406\uff0c\u5c0f\u7684\u65f6\u5019\uff0c\u4f60\u4e5f\u8bb8\u5b8c\u5168\u4e0d\u5728\u610f\uff0c\u8fd9\u5230\u5e95\u662f\u591a\u5c11\u4e2a\u6beb\u79d2\u641e\u5b9a\u7684\u3002\u4f46\u662f\uff0c\u5982\u679c\u4f60\u73b0\u5728\u6709\u4e0a\u767eT\u7684\u6570\u636e\u8981\u5904\u7406\uff0c\u624b\u4e00\u6296\uff0c\u5199\u4e2a\u4e0d\u90a3\u4e48\u9ad8\u6548\u7684\u7b97\u6cd5\uff0c\u591a\u5904\u7406\u4e2a\u51e0\u5929\uff0c\u751a\u81f3\u51e0\u5468\u90fd\u6709\u53ef\u80fd\u3002\u5f53\u7136\uff0c\u66f4\u53ef\u6015\u7684\u5c31\u4e0d\u662f\u6162\uff0c\u800c\u662f\u574f\u4e86 \u2014\u2014 \u4e00\u4e2a\u5c0fbug\u53ef\u4ee5\u574f\u4e86\u6574\u6761pipeline\u3002\u6709\u7684\u65f6\u5019\uff0c\u4f60\u751a\u81f3\u90fd\u4e0d\u80fd\u8bf4\u662fbug\uff0c\u4f46\u662f\u53cd\u6b63\u4e0d\u723d\u5c31\u662f\u4e86\u3002\u6bd4\u5982\uff0c\u4e24\u4e2a\u4eba\u5199\u9884\u5904\u7406\uff0c\u4e00\u4e2a\u4eba\u628a\u56fe\u7247\u5f04\u6210\u4e86BGR\uff0c\u4e00\u4e2a\u5f04\u6210\u4e86RGB \u2014\u2014 \u53c8\u4e0d\u662f\u4e0d\u80fd\u7528\uff0c\u4f46\u662f\u5c31\u662f\u8188\u5e94\u4eba\uff1b\u53c8\u6bd4\u5982\uff0c\u6570\u636e\u539f\u56fe\u592a\u5927\u4e86\uff0c\u8981\u7edf\u4e00\u7f29\u653e\uff0c\u7136\u540e\u6709\u4eba\u505a\u7684\u65f6\u5019\u76f4\u63a5\u5c31\u7f29\u6210\u4e86\u65b9\u5f62\uff0c\u7136\u540e\u5462\uff1f\u6211\u4eec\u4e4b\u540e\u9700\u8981\u7684\u6a21\u578b\u8981\u6b63\u5e38\u957f\u5bbd\u6bd4\u7684\u6570\u636e\u53c8\u8be5\u600e\u4e48\u529e\u5462\uff1f\u518d\u6765\u4e00\u6b21\u561b\uff1f\u518d\u641e\u4e24\u4e2a\u793c\u62dc\uff1f\u4f60\u8bf4\u8fd9\u662f\u4e2a\u4ec0\u4e48\u95ee\u9898\uff1f\u53ef\u4ee5\u770b\u6210\u662f\u79d1\u5b66\uff0c\u5f53\u7136\u4e5f\u53ef\u4ee5\u770b\u6210\u662f\u5de5\u7a0b\u7684\u4e00\u90e8\u5206\uff0c\u8fd9\u4e24\u4e2a\u5c31\u662f\u7d27\u5bc6\u7ed3\u5408\u5728\u4e00\u8d77\u7684\uff0c\u5355\u5355\u4f60\u4f1a\u8c03\u6a21\u578b\uff0c\u5728\u8fd9\u4e2a\u5343\u5361\u8bad\u7ec3\u7684\u4e8b\u60c5\u4e0a\uff0c\u4f60\u662f\u73a9\u4e0d\u8f6c\u7684\u3002\u56e0\u4e3a\u5f88\u591a\u95ee\u9898\u5361\u8116\u5b50\u7684\u95ee\u9898\uff0c\u6839\u672c\u5c31\u662f\uff0c\u6216\u8005\u5927\u6982\u7387\u662f\u5de5\u7a0b\u95ee\u9898 \u2014\u2014 \u4ec0\u4e48\u7f51\u7edc\u901a\u4fe1\uff0c\u4ec0\u4e48\u78c1\u76d8\u7a7a\u95f4\uff0c\u4ec0\u4e48\u8bfb\u53d6\u901f\u5ea6\uff0c\u4ec0\u4e48\u6570\u503c\u7a33\u5b9a\uff0c\u5c0f\u89c4\u6a21\u7684\u65f6\u5019\uff0c\u4f60\u90fd\u53ef\u4ee5\u4e0d\u7528\u7ba1\uff0c\u60f3\u600e\u4e48\u641e\u600e\u4e48\u641e\uff0c\u600e\u4e48\u641e\u53ef\u4ee5\u600e\u4e48\u6709\u3002\u53ef\u662f\u4e0a\u4e86\u89c4\u6a21\u4e4b\u540e\uff0c\u5f88\u591a\u4e1c\u897f\u90fd\u88ab\u9650\u5b9a\u6b7b\u4e86\uff0c\u6839\u672c\u4e0d\u662f\u4f60\u60f3\u600e\u4e48\u5e72\u5c31\u80fd\u600e\u4e48\u5e72\u7684\u3002\u6211\u8bf4\u7684\u8fd9\u4e2a\u8bdd\uff0c\u5927\u6982\u5f88\u591a\u7528pytorch+cuda\u7684\u670b\u53cb\u4e5f\u4e0d\u89c1\u7684\u8ba4\u540c\uff0c\u6bd5\u7adf\u8fd9\u5957\u7ec4\u5408\u4e0b\uff0c\u6ca1\u6709\u592a\u591a\u6280\u672f\u652f\u6301\u7684\u56e2\u961f\u4e5f\u5e72\u6210\u4e86\u8fd9\u6837\u7684\u4e8b\u60c5\u3002\u4f46\u662f\uff0c\u8fd9\u80cc\u540e\u662f\u56e0\u4e3a\uff0c\u4f7f\u7528\u80fd\u652f\u6301\u8fd9\u6837\u8bad\u7ec3\u7684\u4e91\u670d\u52a1\u672c\u8eab\u5c31\u610f\u5473\u7740\uff0c\u4ed8\u4e86\u66f4\u591a\u7684\u94b1\uff0c\u5728\u5df2\u7ecf\u767d\u5ad6\u4e86nvidia\u4e00\u6ce2\u7684\u524d\u63d0\u4e0b\uff0c\u5916\u52a0meta\uff08pytorch\uff09\uff0c\u5916\u52a0\u5fae\u8f6f\uff08deepspeed\uff09\uff0c\u5916\u52a0\u2026\u2026\uff0c\u53c8\u53d8\u76f8\u96c7\u4f63\u4e86\u4e00\u4e2a\u4e13\u95e8\u7684\u652f\u6301\u56e2\u961f\u3002\u4f46\u662f\uff0c\u8fd9\u4e9b\u90fd\u4e0d\u6539\u53d8\u4e00\u4e2a\u4e8b\u5b9e \u2014\u2014 \u90a3\u5c31\u662f\uff0c\u8fd9\u90fd\u662f\u4f60\u8ddf\u7740\u524d\u4eba\u7684\u811a\u6b65\u524d\u8fdb\uff0c\u6709\u4eba\u66ff\u4f60\u5df2\u7ecf\u628a\u8fd9\u6761\u8def\u4e0a\u7684\u5751\uff0c\u8e29\u7684\u5dee\u4e0d\u591a\u4e86\u3002\u53ef\u662f\uff0c\u5982\u679c\u4f60\u8981\u505a\u4e9b\u539f\u521b\u6027\u7684\u5de5\u4f5c\u5462\uff1f\u5fc5\u7136\u662f\u4f1a\u9047\u5230\u5f88\u591a\u524d\u4eba\u90fd\u4e0d\u4f1a\u6709\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u591a\u8bf4\u4e00\u53e5\uff0c\u4e5f\u8bb8\u6709\u4eba\u4f1a\u8bf4\uff0c\u201c\u6211\u4e0d\u5173\u5fc3\u522b\u7684\uff0c\u6211\u5c31\u53ea\u5173\u5fc3pytorch+cuda\u4e0b\uff0c\u505a\u8bad\u7ec3\u7684\u7ecf\u9a8c\u201d\u3002\u90a3\u6211\u544a\u8bc9\u4f60\uff0c\u8fd9\u672c\u8d28\u4e0a\u8fd9\u8ddf\u4f60\u5355\u673a\u5355\u5361\u8bad\u7ec3\u5c31\u4e0d\u5e94\u8be5\u6709\u4ec0\u4e48\u4e0d\u4e00\u6837\uff0c\u8ddf\u662f\u4e0d\u662fpytorch\uff0c\u7528\u4e0d\u7528cuda\u90fd\u6ca1\u4ec0\u4e48\u5927\u5173\u7cfb \u2014\u2014 \u4f60\u60f3\u60f3\u6700\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u8fd9\u662f\u4e0d\u662f\u5c31\u5e94\u8be5\u8ddf\u5355\u673a\u5355\u5361\u8bad\u7ec3\u4e00\u6837\u4e48\uff0c\u65e0\u975e\u5c31\u662f\u73b0\u5728\u7684\u8fd9\u4e2a\u201c\u5355\u673a\u201d\u7684GPU\u5185\u5b58\u662f\u6240\u6709\u7684\u673a\u5668GPU\u5185\u5b58\u7684\u603b\u548c\uff0c\u80fd\u8ba9\u4f60\u7528\u4e00\u4e2a\u66f4\u5927\u7684batch size\u548c\u5b66\u4e60\u7387\u3002\u81f3\u4e8e\uff0cGPU\u5185\u90e8\u600e\u4e48\u901a\u4fe1\uff0c\u6570\u636e\u600e\u4e48\u901a\u4fe1\uff0c\u5404\u4e2a\u673a\u5668\u600e\u4e48\u901a\u4fe1\uff0cgradient\u4f20\u64ad\u600e\u4e48\u5b9e\u73b0\uff0c\u9700\u8981\u4f60\u8fd9\u4e2a\u8bad\u6a21\u5e08\u77e5\u9053\u5417\uff1f\u4f60\u5728\u5355\u673a\u5355\u5361\u7684\u65f6\u5019\u90fd\u4e0d\u7528\u77e5\u9053\uff0c\u5728\u5355\u673a\u591a\u5361\u7684\u65f6\u5019\u4e0d\u7528\u77e5\u9053\uff0c\u5728\u5c0f\u89c4\u6a21\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u65f6\u5019\u4e0d\u7528\u77e5\u9053\uff0c\u90a3\u4e3a\u4ec0\u4e48\u5230\u4e86\u5343\u5361\u7684\u65f6\u5019\uff0c\u4f60\u5c31\u5e94\u8be5\u77e5\u9053\u4e86\uff1f\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u5c31\u7b97\u5230\u4e86\u767e\u4e07\u5361\uff0c\u4e5f\u4e0d\u7528\u505a\u5efa\u6a21\u7684\u4f60\u53bb\u77e5\u9053\u8fd9\u91cc\u7684\u5404\u79cd\u5de5\u7a0b\u5b9e\u8df5\u3002<\/p>\n\n\n\n<p>\u90a3\u5343\u5361\u8bad\u7ec3\u5230\u5e95\u96be\u5728\u54ea\u91cc\u4e86\uff1f\u9996\u5148\uff0c\u5c31\u662f\u96be\u5728\u4e4b\u524d\u63d0\u53ca\u7684\u5de5\u7a0b\u4e0a\u9762\u4e86 \u2014\u2014&nbsp;<strong>\u7b80\u5355\u5047\u8bbe\u4e00\u4e2a\u5361\u5728\u4e00\u5929\u5185\u4e0d\u6302\u6389\u7684\u6982\u7387\u662fp\uff0c\u90a3\u4e48\u73b0\u5728\u73b0\u5728\u5343\u5361\u540c\u65f6\u4e00\u5929\u5185\u4e0d\u6302\u6389\u7684\u6982\u7387\u662f\u591a\u5c11\uff1f<\/strong>\u7b97\u7b97\u4f60\u5c31\u77e5\u9053\uff0c\u5bf9\u4e8ep^1000\uff0c\u5176\u5b9e\u6709\u673a\u5668\u6302\u6389\u624d\u662f\u6b63\u5e38\u7684\u3002\u5982\u679c\u662f\u4e07\u5361\u5462\uff1f\u4f60\u53ef\u89c1\u7684\u662f\uff0c\u673a\u5668N\u8d8a\u591a\uff0cp^N\u5c31\u8d8a\u5c0f\uff0c\u8fd9\u4e2a\u4e8b\u60c5\u5c31\u662f\u8d8a\u96be\u3002\u6709\u4eba\u8981\u8bf4\u201c\u6211\u5355\u673a\u8bad\u7ec3\u7684\u65f6\u5019\uff0c\u51e0\u5e74\u90fd\u9047\u4e0d\u5230\u95ee\u9898\uff0c\u8001\u9ec4\u7684GPU\u7a33\u5b9a\u7684\u4e00\u584c\u7cca\u6d82\u3002\u201d\u5bf9\u6b64\uff0c\u6211\u4e5f\u53ea\u6709\u5475\u5475\uff0c\u6ca1\u4eba\u544a\u8bc9\u4f60\u8bad\u7ec3\u4e0d\u4e0b\u53bb\u90fd\u662fGPU\u7684\u95ee\u9898\u3002\u4f60\u5927\u6982\u662f\u9009\u62e9\u6027\u5fd8\u8bb0\u4e86\u5404\u79cd\u81ea\u5df1\u8bad\u7ec3\u4e2d\u9047\u5230\u7684\u4e8b\u60c5 \u2014\u2014 \u6bd4\u5982\uff0c\u4e0a\u6b21\u5b9e\u9a8c\u4e2d\u65ad\uff0cGPU\u8fdb\u7a0b\u6ca1\u6740\u5e72\u51c0\uff0c\u8fd8\u5360\u7740\u5185\u5b58\uff1b\u548c\u4eba\u5171\u4eab\u7684\u670d\u52a1\u5668\u4e0a\uff0c\u6709\u4e2a\u5367\u9f99\u89c9\u5f97\u4f60\u8bad\u7ec3\u7684\u65f6\u5019CPU\u5360\u7528\u7387\u4f4e\u4e86\u70b9\uff0c\u7ed9\u4f60\u52a0\u4e86\u70b9\u4efb\u52a1\uff1b\u597d\u5de7\u4e0d\u5de7\uff0c\u9ed8\u8ba4\u7684\u7f13\u5b58\u5730\u5740\u6ca1\u5730\u65b9\u4e86\uff0c\u5305\u88c5\u4e0d\u4e0a\u4e86\uff0c\u9884\u8bad\u7ec3\u6a21\u578b\u4e0b\u4e0d\u6765\u4e86\u2026\u2026 \u8bf4\u8fd9\u4e48\u591a\u770b\u4f3c\u548c\u8bad\u7ec3\u65e0\u5173\u7684\u4e8b\u60c5\u662f\u56e0\u4e3a\uff0c\u6240\u6709\u8fd9\u4e9b\u90fd\u8981\u80fd\u81ea\u52a8\u5316\uff0c\u56e0\u4e3a\u91cc\u9762\u6709\u4e00\u4e2a\u5730\u65b9\u7ffb\u8f66\u4e86\uff0c\u4f60\u8bad\u7ec3\u5c31\u8fdb\u884c\u4e0d\u4e0b\u53bb\u4e86\u3002\u901a\u4fe1\u8fde\u4e0d\u4e0a\uff0c\u78c1\u76d8\u6ee1\u4e86\uff0c\u9047\u4e0a\u4e86\u5947\u8469\u7684GPU\uff0cIO\u5de8\u6162 \u2026\u2026 \u4e0d\u7ba1\u4ec0\u4e48\u539f\u56e0\u6302\u6389\u4e86\uff0c\u5173\u952e\u7684\u662f\u4e4b\u540e\u5e94\u8be5\u600e\u4e48\u529e\uff1f\u6709\u6ca1\u6709\u53ef\u80fd\u5bf9\u8fd9\u4e2a\u51fa\u95ee\u9898\u7684\u673a\u5668\u8fdb\u884c\u70ed\u66ff\u6362\uff1f\u600e\u4e48\u529e\u624d\u80fd\u6700\u5927\u7a0b\u5ea6\u4e0d\u5f71\u54cd\u8bad\u7ec3\u8fdb\u7a0b\uff1f\u600e\u4e48\u6837\u624d\u80fd\u5728\u4e0b\u6b21\u907f\u514d\u540c\u6837\u7684\u95ee\u9898\u3002\u5f53\u7136\uff0c\u5b9e\u9645\u60c5\u51b5\u53ef\u4ee5\u66f4\u52a0\u590d\u6742\uff0cGPU\u4e0d\u89c1\u7684\u662f\u540c\u6279\u6b21\u7684\uff0c\u6a21\u578b\u4e5f\u53ef\u4ee5\u5927\u5230\uff0c\u54ea\u6015\u5728A100\u4e0a\u4e5f\u53ea\u80fd\u8fd9\u4e2a\u673a\u5668\u653e\u4e00\u90e8\u5206\uff0c\u90a3\u4e2a\u673a\u5668\u653e\u4e00\u90e8\u5206\u2026\u2026<\/p>\n\n\n\n<p>\u4f46\u662f\u4e5f\u522b\u8bef\u89e3\uff0c\u4ee5\u4e3a\u5343\u5361\u8bad\u7ec3\uff0c\u5c31\u5bf9\u8bad\u6a21\u5e08\u800c\u8a00\uff0c\u5176\u5b9e\u6ca1\u4ec0\u4e48\u6311\u6218\u3002\u8fd9\u6837\u7684\u7406\u89e3\u663e\u7136\u662f\u9519\u7684\u3002\u8fd9\u5bf9\u8bad\u6a21\u5e08\u7684\u5b9e\u64cd\u6765\u8bf4\uff0c\u80af\u5b9a\u662f\u4e00\u4e2a\u5de8\u5927\u7684\u6311\u6218\u3002\u5b8c\u5168\u662f\u62ff\u7740\u5356\u767d\u83dc\u94b1\uff08\u60f3\u60f3\u4f60\u5e74\u85aa\u624d\u591a\u5c11\uff0c\u7b97\u4f60\u5e74\u85aa\u767e\u4e07\u597d\u4e86\uff09\uff0c\u64cd\u7740\u5356\u767d\u7c89\u7684\u5fc3\uff08\u8fd9\u5343\u5361\u8bad\u7ec3\u8981\u82b1\u591a\u5c11\u94b1\uff1f\u4f60\u5e74\u85aa\u90fd\u4e0d\u591f\u5b83\u7684\u4e00\u4e2a\u96f6\u5934\uff09\u3002\u56e0\u4e3a\u8fd9\u673a\u5668\u4e00\u5f00\uff0c\u5b9e\u5728\u662f\u592a\u70e7\u91d1\u4e86\u3002\u800c\u4e14\u53ef\u89c1\u7684\u662f\uff0c\u4f60\u5fc5\u7136\u662f\u8981\u53bbdebug\u7684 \u2014\u2014 \u4e3a\u4ec0\u4e48\u5c0f\u6a21\u578b\u7684\u65f6\u5019\uff0c\u8bad\u7ec3\u7684\u633a\u597d\u7684\uff0c\u4e00\u53d8\u5927\u5c31\u7ffb\u8f66\u4e86\uff1f\u6216\u8005\u8bf4\uff0c\u867d\u7136\u6ca1\u7ffb\u8f66\uff0c\u4f46\u662f\u4e3a\u4ec0\u4e48\u6027\u80fd\u5c31\u6da8\u4e86\u4e00\u4e22\u4e22\uff1f\u6216\u8005\u4e3a\u4ec0\u4e48\u524d\u9762\u8bad\u7ec3\u633a\u7a33\u5b9a\u7684\uff0c\u5230\u4e86\u540e\u9762\u7684loss curve\u5c31\u4f1a\u6709\u5f88\u5927\u7684spike\uff1f\u6709\u7ecf\u9a8c\u7684\u8bad\u6a21\u5e08\u80fd\u66f4\u65e9\uff0c\u66f4\u5feb\u7684\u53d1\u73b0\u95ee\u9898\u3002\u4e5f\u80fd\u66f4\u5feb\u548c\u66f4\u597d\u7684\u89e3\u51b3\u95ee\u9898\u3002\u5f88\u591a\u65f6\u5019\uff0c\u4e5f\u771f\u4e0d\u89c1\u7684\u770blog\u5c31\u80fd\u770b\u51fa\u6765\u70b9\u5565\u7684\uff0c\u770b\u6570\u636e\uff0c\u770bgradient\u7684\u5927\u5c0f\u5206\u5e03\uff0c\u548c\u5176\u4ed6\u6a21\u578b\u7684\u8bad\u7ec3\u8fdb\u884c\u8bb0\u5f55\u505a\u6bd4\u5bf9\uff0c\u751a\u81f3\u505a\u53ef\u89c6\u5316\uff0c\u90fd\u662f\u5f88\u6709\u5fc5\u8981\u7684\u3002\u800c\u8fd9\u6240\u6709\u7684\u4e00\u5207\uff0c\u90fd\u9700\u8981\u4f60\u5f88\u6709\u7ecf\u9a8c \u2014\u2014 \u540c\u6837\u7684log\uff0c\u6709\u4eba\u5c31\u80fd\u4e00\u773c\u770b\u51fa\u6765\u95ee\u9898\u5728\u54ea\u91cc\uff0c\u6709\u4eba\u5c31\u53ea\u80fd\u5bf9\u7740\u53d1\u5446\uff0c\u6216\u8005\u673a\u68b0\u6027\u7684\u8bf4\u201c\u6362\u4e00\u7ec4\u53c2\u6570\u518d\u8bd5\u4e00\u4e0b\u201d\u3002\u540c\u6837\u89c9\u5f97\u53ef\u80fd\u54ea\u91cc\u6709\u95ee\u9898\uff0c\u6709\u4eba\u5c31\u80fd\u77e5\u9053\u5e94\u8be5\u6765\u9a8c\u8bc1\u8fd9\u4e2a\u731c\u60f3\u662f\u5bf9\u662f\u9519\uff0c\u6709\u4eba\u5c31\u53ea\u80fd\u5929\u9a6c\u884c\u7a7a\u7684\u7ed9\u51fa\u4e00\u5806\uff0c\u8c01\u4e5f\u4e0d\u77e5\u9053\u5bf9\u4e0d\u5bf9\u7684\u539f\u56e0\u3002\u6240\u4ee5\uff0c\u4e00\u65e6\u8fd9\u6761\u8def\u7ebf\u88ab\u6478\u7d22\u51fa\u6765\u4e4b\u540e\uff0c\u5176\u5b9e\u4e5f\u5c31\u6ca1\u4ec0\u4e48\u96be\u5ea6\u4e86 \u2014\u2014 \u6570\u636e\uff0c\u811a\u672c\uff0c\u673a\u5668\u90fd\u5728\u90a3\u91cc\u4e86\uff0c\u6211\u5c31\u95ee\u4f60\uff0c\u6211\u5728\u670d\u52a1\u5668\u4e0arun\u90a3\u6761\u5343\u5361\u8bad\u7ec3\u547d\u4ee4\uff0c\u8ddf\u4f60run\u7684\u80fd\u6709\u4ec0\u4e48\u4e0d\u540c\uff1f<strong>\u6240\u4ee5\uff0c\u771f\u6b63\u7684\u5173\u952e\u4e0d\u662f\u5728\u4e8e\u6709\u6ca1\u6709\u7528\u8fc7\u5343\u5361GPU\u8bad\u7ec3\u8fc7\u6a21\u578b\uff0c\u800c\u662f\u6709\u6ca1\u6709\u4ece\u5934\u81f3\u5c3e\uff0c\u4e00\u8def\u62ab\u8346\u65a9\u68d8\u7684\u81ea\u5df1\u6dcc\u51fa\u6765\u4e00\u6761\u53ef\u91cd\u590d\u7684\u6a21\u578b\u6280\u672f\u8def\u7ebf\uff01\uff01<\/strong><\/p>\n\n\n\n<p>\u5f53\u7136\uff0c\u5982\u679c\u4f60\u8981\u4ee5\u4e3a\uff0c\u8fd9\u4e8b\u60c5\u5c31\u53ea\u662f\u6280\u672f\uff0c\u90a3\u4e5f\u662f\u592a\u5e74\u8f7b\u4e86\u70b9\u3002\u673a\u5668\u4e00\u5f00\uff0c\u8981\u591a\u5c11\u94b1\uff0c\u8fd9\u8d26\u771f\u8981\u7b97\u51c6\u4ece\u8bad\u6a21\u5e08\u7684\u89d2\u5ea6\u8bf4\u662f\u5f88\u96be\u7684\uff0c\u6bd5\u7adf\u5177\u4f53\u4ef7\u683c\u90fd\u662f\u5927\u516c\u53f8\u4e4b\u95f4\u534f\u8bae\u7684\uff0c\u5c5e\u4e8e\u5546\u4e1a\u673a\u5bc6\uff0c\u4f46\u662f\u4f30\u7b97\u4e2a\u5927\u6982\u7684\u6570\u76ee\u4e0d\u96be\u3002\u6309\u7167aws\u7684p4d\u7b97\uff088\u5361A100\uff0c\u89c1\u4e0b\u56fe\uff09\uff0c\u4fbf\u5b9c\u7684\u7b97\u6cd5\uff0c\u5343\u5361\u8bad\u7ec3\u4e00\u4e2a\u6708\uff0c\u9700\u8981\u82b1\u8d39 $11.57\/\u6bcf\u53f0\u5c0f\u65f6*24\u5c0f\u65f6\/\u5929*125\u53f0*30\u5929 = $1,041,340\uff1b\u6309\u7167\u963f\u91cc\u4e91\u7684\u7b97\u6cd5\uff0c\u5355\u5361\u5e74\u8d39\u00a5170533.8\uff0c\u4e5f\u5c31\u662f\u00a519.47\u6bcf\u5c0f\u65f6\uff0c\u4f46\u662f\u7b97\u4e0a\u591a\u5361\u7684\u8d39\u7528\uff0c\u8fd9\u5b9e\u9645\u4e0a\u6bd4\u4e0a\u9762aws\u7684\u4ef7\u683c\u66f4\u8d35\u3002\u5f53\u7136\uff0c\u4f60\u4e5f\u8bb8\u80fd\u7528\u66f4\u4fbf\u5b9c\u7684\u4ef7\u683c\u62ff\u5230\u673a\u5668\uff0c\u6bd4\u5982\u522b\u627e\u8fd9\u4e48\u5927\u7684\u4e91\u670d\u52a1\u5e73\u53f0\uff0c\u627e\u4e2a\u5c0f\u7684\uff0c\u4f46\u662f\u518d\u5c11\u8fd8\u80fd\u5c11\u591a\u5c11\uff0c\u7b97\u62535\u6298\uff0c\u8fd9\u90fd\u662f50\u591a\u4e07\u7f8e\u5200\uff0c350\u591a\u4e07\u4eba\u6c11\u5e01\u4e00\u4e2a\u6708\u3002\u8981\u77e5\u9053\uff0c\u8fd9\u53ef\u662f\u8bad\u7ec3\u4e00\u6b21\u7684\u4ef7\u683c\u54e6\u3002\u4e00\u4e2a\u80fd\u7528\u7684\u6a21\u578b\u80cc\u540e\uff0c\u53ef\u662f5x\uff0c\u751a\u81f310x\u66f4\u591a\u7684\u4e0d\u80fd\u7528\u6a21\u578b\u54e6\uff0c\u6240\u4ee5\u70e7\u4e2a\u51e0\u5343\u4e07\uff0c\u771f\u8ddf\u73a9\u4e00\u6837\u3002<\/p>\n\n\n\n<p>\u6b63\u662f\u56e0\u4e3a\u8fd9\u4e48\u8d35\uff0c\u6240\u4ee5\u4e5f\u540c\u6837\u8868\u660e\u4e86\uff0c\u4e3a\u4ec0\u4e48\u4e00\u5b9a\u8981\u627e\u6709\u7ecf\u9a8c\u7684\u8bad\u6a21\u5e08 \u2014\u2014&nbsp;<strong>\u4f60\u8981\u77e5\u9053\uff0c\u4f60\u81ea\u5df1\u7684\u6bcf\u4e2a\u5b9e\u9a8c\u51b3\u5b9a\uff0c\u90fd\u662f\u53d8\u76f8\u7684\u82b1\u51fa\u53bb\u5341\u51e0\uff0c\u51e0\u5341\u751a\u81f3\u4e0a\u767e\u4e07\u7684\u7f8e\u91d1\u3002<\/strong>\u65e9\u53d1\u73b0\u95ee\u9898\uff0c\u65e9\u505c\u4e0b\u6765\uff1b\u65e9\u89e3\u51b3\u95ee\u9898\uff0c\u65e9\u5f00\u59cb\uff1b\u77e5\u9053\u600e\u4e48\u5077\u61d2\uff0c\u4ec0\u4e48\u6837\u7684ablation study\u53ef\u4ee5\u8df3\u4ec0\u4e48\u5fc5\u987b\u505a\uff0c\u4ec0\u4e48\u65f6\u5019\u53ef\u4ee5\u7528\u5c0f\u6a21\u578b\u66ff\u4ee3\uff0c\u4ec0\u4e48\u65f6\u5019\u53ef\u4ee5\u7528\u4e00\u4e2a\u8001\u7684checkpoint\u6765\u4e2ajump start\uff0c\u4ec0\u4e48\u65f6\u5019\u76f4\u63a5\u767d\u5ad6\u8bba\u6587\u4e0a\u7684\u7ed3\u8bba\u5c31\u884c\u2026\u2026\uff0c\u6240\u6709\u8fd9\u4e9b\u90fd\u548c\u82b1\u591a\u5c11\u94b1\u624d\u80fd\u628a\u8fd9\u4e2a\u4e8b\u60c5\u529e\u4e86\uff0c\u8f93\u51fa\u4e00\u4e2a\u8fbe\u6807\u7684\u6a21\u578b\uff0c\u76f4\u63a5\u76f8\u5173 \u3002\u76f8\u53cd\u7684\uff0c\u4e07\u4e00\u4f60\u8981\u627e\u4e2a\u62c9\u57ae\u7684\u8bad\u6a21\u5e08\uff0c\u524d\u9762\u4e0d\u77e5\u9053\u600e\u4e48\u8ba1\u5212\uff0c\u4ee3\u7801\u4e0d\u77e5\u9053\u600e\u4e48\u9a8c\u8bc1\uff0c\u8bad\u7ec3\u8d77\u6765\u4e86\u4e0d\u61c2\u600e\u4e48\u6709\u6548\u76d1\u63a7\uff0c\u6709\u4e86\u5f02\u5e38\u4e0d\u77e5\u9053\u5982\u4f55\u6392\u9664\uff0c\u2026\u2026\uff0c\u6700\u540e\u90fd\u8981\u9760\u7740\u6a21\u578b\u8bad\u7ec3\u5168\u5b8c\u4e86\u4e4b\u540e\u505aevaluation\u624d\u77e5\u9053\u884c\u4e0d\u884c\u7684\u90a3\u79cd\u3002\u90a3\u4e48\u5c31\u7b97\u9884\u7b97\u5168\u82b1\u5b8c\u4e86\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u8bad\u7ec3\u51fa\u6765\uff0c\u6211\u4e5f\u6ca1\u6709\u4ec0\u4e48\u597d\u5947\u602a\u7684\u3002<\/p>\n\n\n\n<p>\u94fa\u57ab\u8fd9\u4e48\u591a\uff0c\u7ec8\u4e8e\u53ef\u4ee5\u6765\u8c08\u8c08\u6700\u540e\u4e00\u4e2a\u5c42\u9762 \u2014\u2014&nbsp;<strong>\u4eba\u60c5\u4e16\u6545\u4e86\u3002\u4f60\u770b\uff0c\u5343\u5361\u8bad\u7ec3\u8fd9\u4e2a\u4e8b\u60c5\uff0c\u6709\u8fd9\u4e48\u5927\u7684\u98ce\u9669\u7ffb\u8f66\uff0c\u8981\u82b1\u8fd9\u4e48\u591a\u7684\u9884\u7b97\uff0c\u90a3\u4e48\u73b0\u5728\u95ee\u9898\u6765\u4e86\uff0c\u4f60\u8981\u662f\u90e8\u95e8\u9886\u5bfc\uff0c\u4f60\u8ba9\u8c01\u6765\u5e72\u8fd9\u4e2a\u4e8b\u60c5\uff1f<\/strong>\u54e6\uff0c\u4f60\u60f3\u653e\u6743\u7ed9\u4e0b\u9762\u7684\u7ecf\u7406\uff0c\u8ba9\u4ed6\u6765\u627e\u4eba\uff1f\u53c8\u6216\u8005\u627e\u4e2a\u521a\u6765\u7684\u535a\u58eb\uff1f\u627e\u4e2a\u9876\u6821+\u9876\u4f1a\u7684\u535a\u58eb\uff1f\u4e0d\u7ba1\u4f60\u600e\u4e48\u627e\uff0c\u53ef\u95ee\u9898\u662f\uff0c\u4f60\u5c31\u8fd9\u4e48\u4fe1\u5f97\u8fc7\u4ed6\u5417\uff1f\u4f60\u600e\u4e48\u4fdd\u8bc1\u4ed6\uff0c\u80fd\u5e72\u8fd9\u4e2a\uff0c\u80fd\u5e72\u597d\u8fd9\u4e2a\uff0c\u4e0d\u4f1a\u4e2d\u9014\u8dd1\u8def\uff0c\u4e0d\u4f1a\u78e8\u6d0b\u5de5\u2026\u2026 \u8981\u77e5\u9053\uff0c\u8fd9\u6837\u5927\u7684\u9879\u76ee\u548c\u9884\u7b97\uff0c\u5982\u679c\u8981\u771f\u5e72\u584c\u4e86\uff0c\u4e0d\u8bf4\u6574\u4e2a\u90e8\u95e8\u8981\u4e00\u8d77\u5b8c\u86cb\uff0c\u81f3\u5c11\u8fd9\u4e00\u6761\u7ebf\u7684\u4eba\u5458\u5fc5\u7136\u662f\u8981\u62c5\u8d23\u7684\uff0c\u54ea\u6015\u662f\u4e3b\u8981\u9886\u5bfc\u4e5f\u8dd1\u4e0d\u6389\u3002\u6240\u4ee5\u55bd\uff0c<strong>\u5173\u952e\u7684\u5173\u952e\u662f\uff0c\u4f60\u5fc5\u987b\u627e\u81ea\u5df1\u4fe1\u7684\u8fc7\u7684\u4eba\uff0c\u8fd8\u8981\u627e\u786e\u5b9e\u6709\u80fd\u529b\u53ef\u4ee5\u62c5\u5f53\u91cd\u4efb\u7684\u4eba \u2014\u2014 \u771f\u6b63\u6700\u540e\u6765\u5e72\u8fd9\u4e2a\u5343\u5361\u8bad\u7ec3\u7684\u4eba\uff0c\u4e0d\u4f46\u81ea\u5df1\u6280\u672f\u8981\u8fc7\u786c\uff0c\u66f4\u662f\u56e2\u961f\u7684\u4e2d\u575a\u529b\u91cf\uff0c\u81f3\u5c11\u4e5f\u8981\u5f97\u5230\u4e00\u4e24\u4e2a\u5927\u5934\u76ee\u7684\u652f\u6301\uff0c\u800c\u4e14\u8fd8\u8981\u5f97\u5230\u5c0f\u5934\u76ee\u652f\u6301\u3002<\/strong>\u4f60\u518d\u725b\u903c\uff0c\u6ca1\u4fe1\u4efb\u6ca1\u5927\u4f6c\u652f\u6301\uff0c\u8fd9\u4e8b\u60c5\u4e0d\u8bf4\u5b8c\u5168\u4e0d\u53ef\u80fd\uff0c\u4e5f\u662f\u57fa\u672c\u6ca1\u53ef\u80fd\u3002\u4f60\u518d\u725b\u903c\uff0c\u8981\u662f\u771f\u7684\u5c0f\u5934\u76ee\u7ed9\u4f60\u4e0a\u773c\u836f\uff0c\u6bd4\u5982\uff0c\u8ddf\u4e0a\u9762\u5439\u98ce\uff0c\u201c\u597d\u50cf\u770b\u89c1\u4f60\u5728\u770b\u62db\u8058\u7f51\u7ad9\u201d\uff0c\u4f60\u60f3\u5927\u5934\u76ee\u5fc3\u91cc\u4f1a\u4e0d\u4f1a\u6709\u9634\u5f71\uff1f\u6240\u4ee5\uff0c\u522b\u7ed9\u6211\u626f\u4ec0\u4e48\uff0c\u8001\u5b50\u6709\u591a\u5c11\u9876\u4f1a\u9876\u520a\uff0c\u8001\u5b50\u5bfc\u5e08\u662f\u8c01\u8c01\u8c01\uff0c\u8fd9\u5728\u7edd\u5927\u591a\u6570\u60c5\u51b5\uff0c\u90fd\u4e0d\u597d\u4f7f\u3002\u6240\u4ee5\uff0c\u521a\u6bd5\u4e1a\u7684\uff0c\u6216\u8005\u505a\u5b9e\u4e60\u7684\uff0c\u6216\u8005\u521a\u5de5\u4f5c\u7684\uff0c\u5982\u679c\u5ba3\u79f0\u81ea\u5df1\u6709\u8fd9\u4e2a\u7ecf\u9a8c\uff0c\u5c31\u662f\u4e00\u773c\u4e01\u771f\u3002\u56e0\u4e3a\u4e0a\u9762\u662f\u7edd\u5bf9\u4e0d\u4f1a\u627e\u4e0d\u4fe1\u4efb\u7684\u4eba\u6765\u8fd9\u6837\u91cd\u8981\u7684\u5de5\u4f5c\uff0c\u8fd9\u8ddf\u4f60\u6709\u6ca1\u6709\u76f8\u5173\u7684\u5de5\u4f5c\u7ecf\u9a8c\u65e0\u5173\u3002<strong>\u8fd9\u540c\u6837\u610f\u5473\u7740\uff0c\u771f\u6b63\u5e72\u8fd9\u4e9b\u4e8b\u60c5\u7684\u4eba\u4e5f\u5f88\u96be\u6d41\u51fa\u6765 \u2014\u2014 \u56e0\u4e3a\u5bf9\u4e8e\u5ae1\u7cfb\u6765\u8bf4\uff0c\u52a0\u85aa\u5347\u804c\uff0c\u5728\u5e72\u597d\u4e86\u7684\u524d\u63d0\u4e0b\uff0c\u90a3\u8fd8\u4e0d\u90fd\u662fso easy\u5417\uff1f<\/strong>\u6240\u4ee5\uff0c\u662f\u4f60\uff0c\u4f60\u613f\u610f\u51fa\u6765\u5417\uff1f\u51fa\u6765\u4e86\uff0c\u5c31\u7b97\u4f60\u725b\u903c\uff0c\u4f46\u662f\u83b7\u53d6\u4fe1\u4efb\uff0c\u6210\u4e3a\u5ae1\u7cfb\u4e5f\u8981\u4e00\u4e2a\u65f6\u95f4\uff0c\u4e0d\u662f\u5417\uff1f<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e00\u4e2a\u77e5\u4e4e\u95ee\u9898:\u5982\u4f55\u5224\u65ad\u5019\u9009\u4eba\u6709\u6ca1\u6709\u5343\u5361GPU\u96c6\u7fa4\u7684\u8bad\u7ec3\u7ecf\u9a8c\uff1f\u786e\u5b9e\u5bf9\u4e8e\u666e\u901a\u5f00\u53d1\u8005\u6765\u8bf4\uff0c\u5927\u6982\u7387\u4ece\u672a\u5c1d\u8bd5\u8fc7 &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2024\/09\/18\/discuss-how-to-train-a-model-using-1024-gpus\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u8ba8\u8bba\uff1a\u5982\u4f55\u75281024\u5f20\u663e\u5361\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b?<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[3,11,4,39,26,38],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/19021"}],"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=19021"}],"version-history":[{"count":63,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/19021\/revisions"}],"predecessor-version":[{"id":19099,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/19021\/revisions\/19099"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=19021"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=19021"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=19021"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}