{"id":53736,"date":"2025-02-16T13:51:13","date_gmt":"2025-02-16T05:51:13","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53736\/"},"modified":"2025-02-16T13:51:13","modified_gmt":"2025-02-16T05:51:13","slug":"ddp%ef%bc%9a%e5%88%86%e5%b8%83%e5%bc%8f%e6%95%b0%e6%8d%ae%e5%b9%b6%e8%a1%8c","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53736\/","title":{"rendered":"DDP\uff1a\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c"},"content":{"rendered":"<p>\u6700\u8fd1\uff0c\u5728\u4f7f\u7528 OpenAI \u7684\u8bba\u6587\u201c\u8bed\u8a00\u6a21\u578b\u662f\u65e0\u76d1\u7763\u7684\u591a\u4efb\u52a1\u5b66\u4e60\u8005\u201d\u548c Andrej Karpathy \u7684 YouTube \u89c6\u9891\u201c\u8ba9\u6211\u4eec\u91cd\u73b0 GPT-2 (124M)\u201d\u4ece\u5934\u91cd\u73b0 GPT-2 LLM \u65f6\uff0c\u6211\u5f3a\u70c8\u5730\u60f3\u8981\u4e86\u89e3\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u7684\u5de5\u4f5c\u539f\u7406\u3002\u8bad\u7ec3\u5982\u6b64\u5927\u7684\u6a21\u578b\u9700\u8981\u591a GPU \u8bbe\u7f6e\uff0c\u800c\u4e14\u7531\u4e8e\u8fd9\u662f\u6211\u7b2c\u4e00\u6b21\u5c1d\u8bd5\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u8fd9\u79cd\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u6240\u4ee5\u8fd9\u4e2a\u4e3b\u9898\u5bf9\u6211\u6765\u8bf4\u662f\u5168\u65b0\u7684\u3002<\/p>\n<p>\u4e3a\u4e86\u5f25\u8865\u8fd9\u4e00\u77e5\u8bc6\u5dee\u8ddd\uff0c\u6211\u7acb\u5373\u9605\u8bfb\u4e86 PyTorch \u7684 DDP \u6587\u6863\u5e76\u7cfb\u7edf\u5730\u7406\u89e3\u5b83\u3002\u672c\u6587\u5c31\u662f\u8fd9\u6bb5\u5b66\u4e60\u4e4b\u65c5\u7684\u6210\u679c\u3002<\/p>\n<p>\u968f\u7740\u6570\u636e\u96c6\u548c\u6a21\u578b\u53d8\u5f97\u8d8a\u6765\u8d8a\u5927\uff0c\u5728\u591a\u4e2a GPU \u4e0a\u5206\u914d\u5de5\u4f5c\u8d1f\u8f7d\u4e0d\u4ec5\u6709\u7528\uff0c\u800c\u4e14\u5fc5\u4e0d\u53ef\u5c11\u3002\u5b83\u663e\u8457\u51cf\u5c11\u4e86\u8bad\u7ec3\u65f6\u95f4\uff0c\u589e\u5f3a\u4e86\u53ef\u6269\u5c55\u6027\uff0c\u5e76\u4f7f\u8bad\u7ec3\u5927\u89c4\u6a21\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002PyTorch \u7684\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u662f\u6ee1\u8db3\u8fd9\u4e9b\u9700\u6c42\u7684\u5f3a\u5927\u89e3\u51b3\u65b9\u6848\u4e4b\u4e00\u3002<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u5c06\u89e3\u91ca\u6211\u5bf9 DDP \u7684\u7406\u89e3\u3001\u5b83\u76f8\u5bf9\u4e8e\u6570\u636e\u5e76\u884c (DP) \u7684\u4f18\u52bf\u3001\u5b83\u7684\u5185\u90e8\u5de5\u4f5c\u539f\u7406\u4ee5\u53ca\u4f7f\u7528 ToyModel \u7684\u5b9e\u9645\u5b9e\u73b0\u793a\u4f8b\u3002<\/p>\n<p>\u4f7f\u7528 Dall-E \u521b\u5efa\u7684\u6a21\u578b\u5e76\u884c\u6027\u56fe\u50cf<\/p>\n<p> \u662f PyTorch \u4e2d\u7684\u4e00\u4e2a\u5f3a\u5927\u6a21\u5757\uff0c\u5b83\u5141\u8bb8\u6211\u4eec\u5728\u591a\u53f0\u673a\u5668\u4e0a\u5e76\u884c\u5316\u6211\u4eec\u7684\u6a21\u578b\uff0c\u901a\u8fc7\u5728\u591a\u4e2a GPU \u4e0a\u590d\u5236\u6a21\u578b\u6765\u5b9e\u73b0\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002\u6bcf\u4e2a\u8fdb\u7a0b\u5728\u6570\u636e\u5b50\u96c6\u4e0a\u8bad\u7ec3\u6a21\u578b\u7684\u526f\u672c\uff0c\u5e76\u4e14\u68af\u5ea6\u5728\u8fdb\u7a0b\u4e4b\u95f4\u540c\u6b65\u4ee5\u786e\u4fdd\u6a21\u578b\u53c2\u6570\u7684\u4e00\u81f4\u66f4\u65b0\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u5927\u578b\u6a21\u578b\u65f6\uff0c\u5355\u4e2a GPU \u901a\u5e38\u7f3a\u4e4f\u6709\u6548\u5904\u7406\u4efb\u52a1\u6240\u9700\u7684\u5185\u5b58\u6216\u8ba1\u7b97\u80fd\u529b\u3002\u5373\u4f7f\u53ef\u4ee5\u5728\u5355\u4e2a GPU \u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6240\u9700\u7684\u65f6\u95f4\u4e5f\u53ef\u80fd\u975e\u5e38\u957f\u3002\u4f7f\u7528\u591a\u4e2a GPU \u53ef\u4ee5\u5e2e\u52a9\u6211\u4eec\uff1a<\/p>\n<ul>\n<li>\u5904\u7406\u66f4\u5927\u7684\u6a21\u578b\uff1a\u5728 GPU \u4e4b\u95f4\u805a\u5408\u5185\u5b58\u4ee5\u8bad\u7ec3\u65e0\u6cd5\u5bb9\u7eb3\u5728\u5355\u4e2a GPU \u5185\u5b58\u4e2d\u7684\u6a21\u578b\u3002<\/li>\n<li>\u51cf\u5c11\u8bad\u7ec3\u65f6\u95f4\uff1a\u5728\u8bbe\u5907\u4e4b\u95f4\u5206\u914d\u8ba1\u7b97\uff0c\u5b9e\u73b0\u66f4\u5feb\u7684\u8fed\u4ee3\u3002<\/li>\n<li>\u5b9e\u73b0\u53ef\u6269\u5c55\u6027\uff1a\u968f\u7740\u6570\u636e\u96c6\u5927\u5c0f\u7684\u589e\u52a0\uff0c\u53ef\u8f7b\u677e\u5c06\u8bad\u7ec3\u6269\u5c55\u5230\u66f4\u591a GPU\u3002<\/li>\n<\/ul>\n<p> DDP \u6d41\u7a0b <\/p>\n<h2>1\u3001\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) vs. \u6570\u636e\u5e76\u884c (DP)<\/h2>\n<p>\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u5728\u6027\u80fd\u548c\u7075\u6d3b\u6027\u65b9\u9762\u5747\u4f18\u4e8e\u6570\u636e\u5e76\u884c (DP)\uff0c\u6709\u6548\u89e3\u51b3\u4e86 DP \u7684\u5c40\u9650\u6027\uff1a<\/p>\n<blockquote><p>\n  \u591a\u8fdb\u7a0b\u67b6\u6784\uff1a\n<\/p><\/blockquote>\n<ul>\n<li>DP \u662f\u5355\u8fdb\u7a0b\u548c\u591a\u7ebf\u7a0b\u7684\uff0c\u4ec5\u5728\u4e00\u53f0\u673a\u5668\u4e0a\u8fd0\u884c\u3002\u8fd9\u79cd\u8bbe\u8ba1\u901a\u5e38\u4f1a\u5bfc\u81f4\u7ebf\u7a0b\u4e4b\u95f4\u7684\u5168\u5c40\u89e3\u91ca\u5668\u9501 (GIL) \u4e89\u7528\uff0c\u4ece\u800c\u51cf\u6162\u8ba1\u7b97\u901f\u5ea6\u3002<\/li>\n<li>DDP \u4f7f\u7528\u591a\u8fdb\u7a0b\u65b9\u6cd5\uff0c\u5176\u4e2d\u6bcf\u4e2a GPU \u7531\u5176\u81ea\u5df1\u7684\u8fdb\u7a0b\u7ba1\u7406\u3002\u8fd9\u6d88\u9664\u4e86 GIL \u4e89\u7528\u5e76\u663e\u8457\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\uff0c\u5373\u4f7f\u5728\u5355\u53f0\u673a\u5668\u4e0a\u4e5f\u662f\u5982\u6b64\u3002<\/li>\n<\/ul>\n<p>\u5168\u5c40\u89e3\u91ca\u5668\u9501 (GIL) \u662f Python \u4e2d\u7684\u4e00\u79cd\u673a\u5236\uff0c\u5373\u4f7f\u5728\u591a\u7ebf\u7a0b\u7a0b\u5e8f\u4e2d\u4e5f\u53ea\u5141\u8bb8\u4e00\u4e2a\u7ebf\u7a0b\u4e00\u6b21\u6267\u884c Python \u5b57\u8282\u7801\u3002\u5f53\u8bb8\u591a\u7ebf\u7a0b\u8bd5\u56fe\u540c\u65f6\u8fd0\u884c Python \u4ee3\u7801\u65f6\uff0c\u8fd9\u53ef\u80fd\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/p>\n<p>GIL \u4e89\u7528\u53d1\u751f\u5728\u591a\u4e2a\u7ebf\u7a0b\u7ade\u4e89 GIL \u65f6\uff0c\u7531\u4e8e\u4e00\u6b21\u53ea\u80fd\u6267\u884c\u4e00\u4e2a\u7ebf\u7a0b\u800c\u5176\u4ed6\u7ebf\u7a0b\u5fc5\u987b\u7b49\u5f85\uff0c\u56e0\u6b64\u4f1a\u5bfc\u81f4\u5ef6\u8fdf\u3002<\/p>\n<blockquote><p>\n  \u8de8\u673a\u5668\u53ef\u6269\u5c55\u6027\uff1a\n<\/p><\/blockquote>\n<ul>\n<li>\u7531\u4e8e DP \u4ec5\u9650\u4e8e\u5355\u53f0\u673a\u5668\uff0c\u56e0\u6b64\u4e0d\u9002\u5408\u5927\u89c4\u6a21\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002<\/li>\n<li>DDP \u652f\u6301\u5355\u673a\u548c\u591a\u673a\u8bbe\u7f6e\uff0c\u53ef\u5728\u96c6\u7fa4\u4e2d\u7684\u591a\u4e2a\u8282\u70b9\u4e0a\u65e0\u7f1d\u6269\u5c55\u5927\u578b\u6570\u636e\u96c6\u548c\u6a21\u578b\u3002<\/li>\n<\/ul>\n<p>\u4e0a\u9762\u7684\u6570\u636e\u5e76\u884c (DP) \u4e2d\u7684\u5355\u673a\u662f\u6307\u4ec5\u9650\u4e8e\u4e00\u53f0\u673a\u5668\uff0c\u5373 DP \u53ea\u80fd\u4f7f\u7528\u4e00\u53f0\u8ba1\u7b97\u673a\u4e0a\u53ef\u7528\u7684 GPU\u3002\u5982\u679c\u4f60\u7684\u8bad\u7ec3\u8bbe\u7f6e\u9700\u8981\u7684 GPU \u8d85\u8fc7\u5355\u53f0\u8ba1\u7b97\u673a\u53ef\u4ee5\u63d0\u4f9b\u7684 GPU\uff0c\u5219 DP \u65e0\u6cd5\u4f7f\u7528\u7f51\u7edc\u4e2d\u5176\u4ed6\u673a\u5668\u7684 GPU\u3002<\/p>\n<p>\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\u5047\u8bbe\u4f60\u6709\u4e00\u4e2a\u5305\u542b 4 \u53f0\u673a\u5668\u7684\u96c6\u7fa4\uff0c\u6bcf\u53f0\u673a\u5668\u6709 4 \u4e2a GPU\u3002 DP \u53ea\u80fd\u4f7f\u7528\u4e00\u53f0\u673a\u5668\u4e0a\u7684 4 \u4e2a GPU\uff0c\u800c\u5176\u4ed6\u673a\u5668\u4e0a\u7684 GPU \u5219\u95f2\u7f6e\u4e0d\u7528\u3002<\/li>\n<li>\u53e6\u4e00\u65b9\u9762\uff0c\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u53ef\u4ee5\u4f7f\u7528\u96c6\u7fa4\u4e2d\u6240\u6709\u673a\u5668\u7684 GPU\uff0c\u4f7f\u5176\u5728\u5927\u89c4\u6a21\u8bad\u7ec3\u4e2d\u66f4\u5177\u53ef\u6269\u5c55\u6027\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u4e0e\u6a21\u578b\u5e76\u884c\u7684\u517c\u5bb9\u6027\uff1a\n<\/p><\/blockquote>\n<ul>\n<li>\u5f53\u6a21\u578b\u592a\u5927\u800c\u65e0\u6cd5\u653e\u5728\u5355\u4e2a GPU \u4e0a\u65f6\uff0c\u6a21\u578b\u5e76\u884c\u7528\u4e8e\u5c06\u6a21\u578b\u62c6\u5206\u5230\u591a\u4e2a GPU \u4e0a\u3002DP \u76ee\u524d\u4e0d\u652f\u6301\u5c06\u6a21\u578b\u5e76\u884c\u4e0e\u6570\u636e\u5e76\u884c\u76f8\u7ed3\u5408\u3002<\/li>\n<li>DDP \u4e0e\u6a21\u578b\u5e76\u884c\u65e0\u7f1d\u534f\u4f5c\u3002\u6bcf\u4e2a DDP \u8fdb\u7a0b\u90fd\u53ef\u4ee5\u5728\u5176\u5206\u914d\u7684 GPU \u4e2d\u5229\u7528\u6a21\u578b\u5e76\u884c\uff0c\u5e76\u4e14\u6240\u6709\u8fdb\u7a0b\u5171\u540c\u4f7f\u7528\u6570\u636e\u5e76\u884c\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u6781\u5927\u6a21\u578b\u7684\u9ad8\u6548\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h2>2\u3001\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u7684\u5185\u90e8\u673a\u5236<\/h2>\n<p>\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u662f PyTorch \u4e2d\u7684\u4e00\u4e2a\u5f3a\u5927\u6846\u67b6\uff0c\u65e8\u5728\u6709\u6548\u5730\u5c06\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u8bad\u7ec3\u5206\u5e03\u5728\u591a\u4e2a GPU \u6216\u673a\u5668\u4e0a\u3002\u4ee5\u4e0b\u662f DDP \u5185\u90e8\u8fd0\u4f5c\u65b9\u5f0f\u7684\u8be6\u7ec6\u5206\u6b65\u8bf4\u660e\uff1a<\/p>\n<blockquote><p>\n  \u6a21\u578b\u521d\u59cb\u5316\u548c\u590d\u5236\n<\/p><\/blockquote>\n<p>\u8fdb\u7a0b\u7ec4\u8bbe\u7f6e\uff1aDDP \u4f9d\u8d56\u4e8e c10d ProcessGroup \u8fdb\u884c\u8fdb\u7a0b\u95f4\u901a\u4fe1\u3002\u5728\u6784\u9020 DDP \u5b9e\u4f8b\u4e4b\u524d\uff0c\u5fc5\u987b\u521d\u59cb\u5316 ProcessGroup \u4ee5\u5efa\u7acb\u8fdb\u7a0b\u4e4b\u95f4\u7684\u901a\u4fe1\u3002\uff08\u2018gloo\u2019\u3001\u2018nccl\u2019\u3001\u2018mpi\u2019\uff09<\/p>\n<p> PyTorch \u9644\u5e26\u7684\u8fdb\u7a0b\u7ec4 <\/p>\n<ul>\n<li>\u6a21\u578b\u5e7f\u64ad\uff1a\u5728\u521d\u59cb\u5316\u671f\u95f4\uff0c\u6a21\u578b\u7684 state_dict \u4ece\u7b49\u7ea7\u4e3a 0 \u7684\u8fdb\u7a0b\u5e7f\u64ad\u5230\u6240\u6709\u5176\u4ed6\u8fdb\u7a0b\u3002\u8fd9\u53ef\u786e\u4fdd\u6a21\u578b\u7684\u6240\u6709\u526f\u672c\u90fd\u4ee5\u76f8\u540c\u7684\u53c2\u6570\u5f00\u59cb\u3002<\/li>\n<li>Reducer \u521b\u5efa\uff1a\u6bcf\u4e2a DDP \u8fdb\u7a0b\u90fd\u4f1a\u521d\u59cb\u5316\u4e00\u4e2a Reducer\uff0c\u8d1f\u8d23\u7ba1\u7406\u68af\u5ea6\u540c\u6b65\u3002Reducer \u5c06\u68af\u5ea6\u7ec4\u7ec7\u5230\u5b58\u50a8\u6876\u4e2d\u4ee5\u4f18\u5316\u901a\u4fe1\u3002\u53ef\u4ee5\u4f7f\u7528 DDP \u6784\u9020\u51fd\u6570\u4e2d\u7684 bucket_cap_mb \u53c2\u6570\u914d\u7f6e\u5b58\u50a8\u6876\u5927\u5c0f\u3002<\/li>\n<\/ul>\n<p> \u4f7f\u7528\u54ea\u4e2a\u8fdb\u7a0b\u7ec4 <\/p>\n<blockquote><p>\n  \u524d\u5411\u4f20\u9012\n<\/p><\/blockquote>\n<ul>\n<li>\u6bcf\u4e2a GPU \u4f7f\u7528\u5176\u6a21\u578b\u7684\u672c\u5730\u526f\u672c\u72ec\u7acb\u5904\u7406\u5176\u5c0f\u6279\u91cf\u3002<\/li>\n<li>\u5982\u679c\u8bbe\u7f6e\u4e86 find_unused_pa\u200b\u200brameters=True\uff0cDDP \u4f1a\u904d\u5386\u81ea\u52a8\u6c42\u5bfc\u56fe\u4ee5\u8bc6\u522b\u4e0d\u9700\u8981\u68af\u5ea6\u8ba1\u7b97\u7684\u53c2\u6570\u3002\u8fd9\u53ef\u786e\u4fdd DDP \u4ec5\u5728\u53cd\u5411\u4f20\u9012\u671f\u95f4\u540c\u6b65\u6d3b\u52a8\u53c2\u6570\u7684\u68af\u5ea6\u3002<\/li>\n<li>\u6ce8\u610f\uff1a\u904d\u5386\u56fe\u4f1a\u5e26\u6765\u5f00\u9500\uff0c\u56e0\u6b64\u5efa\u8bae\u4ec5\u5728\u5fc5\u8981\u65f6\u542f\u7528 find_unused_pa\u200b\u200brameters\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u53cd\u5411\u4f20\u9012\u548c\u68af\u5ea6\u540c\u6b65\n<\/p><\/blockquote>\n<ul>\n<li>\u81ea\u52a8\u6c42\u5bfc\u94a9\u5b50\uff1aDDP \u5728\u521d\u59cb\u5316\u671f\u95f4\u6ce8\u518c\u94a9\u5b50\u4ee5\u540c\u6b65\u5728\u53cd\u5411\u4f20\u9012\u671f\u95f4\u53ef\u7528\u7684\u68af\u5ea6\u3002<\/li>\n<li>\u6876\u5f0f\u68af\u5ea6\u51cf\u5c11\uff1a<\/li>\n<li>\u68af\u5ea6\u88ab\u5206\u7ec4\u5230\u6876\u4e2d\uff08\u57fa\u4e8e\u6a21\u578b\u53c2\u6570\u987a\u5e8f\uff09\u4ee5\u4f18\u5316\u901a\u4fe1\u3002<\/li>\n<\/ul>\n<p> \u5206\u914d\u7ed9\u4e0d\u540c\u6876\u7684\u68af\u5ea6 <\/p>\n<ul>\n<li>\u5f53\u6876\u4e2d\u7684\u6240\u6709\u68af\u5ea6\u90fd\u51c6\u5907\u5c31\u7eea\u65f6\uff0cDDP \u6267\u884c\u5f02\u6b65 allreduce \u64cd\u4f5c\u4ee5\u8ba1\u7b97\u6240\u6709\u8fdb\u7a0b\u7684\u5e73\u5747\u68af\u5ea6\u3002<\/li>\n<li>\u4e00\u65e6\u6240\u6709 allreduce \u64cd\u4f5c\u5b8c\u6210\uff0c\u5e73\u5747\u68af\u5ea6\u5c31\u4f1a\u5199\u56de\u5230\u76f8\u5e94\u7684\u53c2\u6570\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u4f18\u5316\u5668\u6b65\u9aa4\n<\/p><\/blockquote>\n<ul>\n<li>\u540c\u6b65\u540e\uff0c\u4f18\u5316\u5668\u4f7f\u7528\u5e73\u5747\u68af\u5ea6\u66f4\u65b0\u6bcf\u4e2a\u672c\u5730\u6a21\u578b\u526f\u672c\u7684\u53c2\u6570\u3002<\/li>\n<li>\u7531\u4e8e\u6240\u6709\u526f\u672c\u90fd\u4ee5\u76f8\u540c\u7684\u53c2\u6570\u5f00\u59cb\u5e76\u63a5\u6536\u76f8\u540c\u7684\u68af\u5ea6\u66f4\u65b0\uff0c\u56e0\u6b64\u5b83\u4eec\u7684\u72b6\u6001\u5728\u5404\u4e2a\u8fdb\u7a0b\u4e4b\u95f4\u4fdd\u6301\u540c\u6b65\u3002<\/li>\n<\/ul>\n<p> DDP \u5185\u90e8\u8bbe\u8ba1 <\/p>\n<h2>3\u3001\u4f7f\u7528 DDP \u7684\u73a9\u5177\u7aef\u5230\u7aef\u793a\u4f8b<\/h2>\n<p>\u8ba9\u6211\u4eec\u6765\u770b\u770b\u5728 PyTorch \u4e2d\u4f7f\u7528 DDP \u7684\u4e00\u4e2a\u5b9e\u9645\u793a\u4f8b\uff0c\u6211\u663e\u7136\u662f\u4ece PyTorch \u5b98\u65b9\u6587\u6863\u4e2d\u83b7\u53d6\u7684\uff0c\u5c31\u50cf\u4e0a\u9762\u7684\u535a\u5ba2\u5185\u5bb9\u4e00\u6837\u3002\u8fd9\u4e2a\u73a9\u5177\u793a\u4f8b\u5728\u591a\u4e2a GPU \u4e0a\u5bf9\u968f\u673a\u6570\u636e\u8bad\u7ec3\u4e00\u4e2a\u7b80\u5355\u7684\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<pre><code>import os\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.utils.data import DataLoader, Dataset\nimport torch.distributed as dist\nfrom torch.nn.parallel import DistributedDataParallel as DDP\n\n\n# Initializes a distributed environment for training.\n# # On Windows platform, the torch.distributed package only\n# supports Gloo backend, FileStore and TcpStore.\n# For FileStore, set init_method parameter in init_process_group\n# to a local file. Example as follow:\n# init_method=\"file:\/\/\/f:\/libtmp\/some_file\"\n# dist.init_process_group(\n#    \"gloo\",\n#    rank=rank,\n#    init_method=init_method,\n#    world_size=world_size)\ndef setup(rank, world_size):\n    torch.distributed.init_process_group(\n        backend='nccl',\n        rank=rank,\n        world_size=world_size\n    )\n\n# Here we have Created a very simple Neural Network\nclass SimpleModel(nn.Module):\n    def __init__(self):\n        super(SimpleModel, self).__init__()\n        self.net1 = nn.Linear(10, 10)\n        self.relu = nn.ReLU()\n        self.net2 = nn.Linear(10, 1)\n\n    def forward(self, x):\n        return self.net2(self.relu(self.net1(x)))\n\n# creatging the sample datset for testing\nclass RandomDataset(Dataset):\n    def __init__(self, size, length):\n        self.data = torch.randn(length, size)\n        self.labels = torch.randn(length, 1)\n\n    def __len__(self):\n        return len(self.data)\n\n    def __getitem__(self, index):\n        return self.data[index], self.labels[index]\n\n# train loop in which setup is done \n# where world_size is the number of GPUs we want to access\n# and rank is the current GPU of interest\ndef train(rank, world_size):\n    setup(rank, world_size)\n    torch.backends.cudnn.benchmark = True  # Optional performance optimization\n\n    model = SimpleModel().to(rank)\n# Converting each model TO DDP object\n    model = DDP(model, device_ids=[rank])\n\n    dataset = RandomDataset(10, 1000)\n    sampler = torch.utils.data.distributed.DistributedSampler(\n        dataset, num_replicas=world_size, rank=rank\n    )\n    dataloader = DataLoader(dataset, batch_size=32, sampler=sampler)\n\n    criterion = nn.MSELoss()\n    optimizer = optim.SGD(model.parameters(), lr=0.01)\n\n    for epoch in range(5):\n        sampler.set_epoch(epoch)  # Ensure proper shuffling\n        for batch, (data, labels) in enumerate(dataloader):\n            data, labels = data.to(rank), labels.to(rank)\n\n            optimizer.zero_grad()\n            outputs = model(data)\n            loss = criterion(outputs, labels)\n            loss.backward()\n            optimizer.step()\n\n            if batch % 10 == 0 and rank == 0:\n                print(f\"Rank {rank}, Epoch {epoch}, Batch {batch}, Loss {loss.item()}\")\n\n    torch.distributed.destroy_process_group()  # Graceful shutdown\n\nif __name__ == '__main__':\n    world_size = torch.cuda.device_count()\n    os.environ['MASTER_ADDR'] = '127.0.0.1'\n    os.environ['MASTER_PORT'] = '29500'\n    torch.multiprocessing.spawn(train, args=(world_size,), nprocs=world_size, join=True)<\/code><\/pre>\n<p>\u5728\u4e0a\u9762\u7684\u4ee3\u7801\u4e2d\uff1a<\/p>\n<p>\u8bbe\u7f6e\uff1a<\/p>\n<ul>\n<li>torch.distributed.init_process_group\uff1a\u521d\u59cb\u5316 GPU \u4e4b\u95f4\u7684\u901a\u4fe1\u3002<\/li>\n<\/ul>\n<p>\u6a21\u578b\u5305\u88c5\uff1a<\/p>\n<ul>\n<li>DistributedDataParallel\uff1a\u5305\u88c5\u6a21\u578b\u4ee5\u5b9e\u73b0\u8de8 GPU \u7684\u540c\u6b65\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<p>DataLoader\uff1a<\/p>\n<ul>\n<li>DistributedSampler\uff1a\u786e\u4fdd\u6bcf\u4e2a GPU \u83b7\u5f97\u4e0d\u540c\u7684\u6570\u636e\u5b50\u96c6\u3002<\/li>\n<\/ul>\n<p>\u8fdb\u7a0b\u751f\u6210\uff1a<\/p>\n<ul>\n<li>torch.multiprocessing.spawn\uff1a\u542f\u52a8\u591a\u4e2a\u8fdb\u7a0b\uff0c\u6bcf\u4e2a\u8fdb\u7a0b\u90fd\u4e0e\u4e00\u4e2a GPU \u7ed1\u5b9a\u3002<\/li>\n<\/ul>\n<h2>5\u3001\u5728\u6a21\u578b\u8bad\u7ec3\u4e2d\u5b9e\u65bd DDP \u7684\u4e00\u4e9b\u5b9e\u9645\u89c2\u5bdf<\/h2>\n<p>\u5982\u679c\u6267\u884c DDP \u5e76\u5177\u6709\u81ea\u5b9a\u4e49\u4f18\u5316\u5668\uff0c\u6216\u8005 DDP \u663e\u793a\u4e0e\u4f18\u5316\u5668\u76f8\u5173\u7684\u9519\u8bef\uff0c\u6700\u597d\u5c06\u539f\u59cb nn.module \u6a21\u578b\u653e\u5728\u4e00\u8fb9\uff0c\u4ee5\u4fbf\u4f18\u5316\u5668\u6267\u884c\u4f18\u5316\uff0c\u5982\u4e0b\u9762\u7684\u4ee3\u7801\u7247\u6bb5\u6240\u793a\u3002\u5176\u4f59\u4f18\u5316\u5668\u76f8\u5173\u4ee3\u7801\u4e0d\u9700\u8981\u66f4\u6539\u3002<\/p>\n<pre><code># Now after setting up DDP it is required and mandatory to wrap the model in DDP\nif ddp:\n    model = DDP(model, device_ids = [ddp_local_rank])\nraw_model = model.module if ddp else model # always contains the \"raw\" unwrapped model\n\n# ========================\n\noptimizer = raw_model.configure_optimizers(weight_decay = 0.1,\n                                       learning_rate = 6e-4,\n                                      device_type = device)<\/code><\/pre>\n<p>\u786e\u4fdd\u6bcf\u4e2a DDP \u8fdb\u7a0b\u6839\u636e\u5176\u8fdb\u7a0b\u6392\u540d\u63a5\u6536\u552f\u4e00\u7684\u8bad\u7ec3\u6570\u636e\u5b50\u96c6\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u4f5c\u4e3a\u53c2\u8003\uff0c\u4e3a\u6bcf\u4e2a DDP \u8fdb\u7a0b\u9002\u5f53\u5206\u914d\u6570\u636e\uff1a<\/p>\n<pre><code>class DataLoaderLite:\n    def __init__(self, B, T, process_rank, num_process):\n        self.B = B\n        self.T = T\n        self.process_rank = process_rank\n        self.num_processes = num_process\n        \n        # at init load tokens from disk and store them in memory\n        # with open('input.txt','r') as f:\n        with open(runpod_absolute_path,'r') as f:\n            text = f.read()\n        enc = tiktoken.get_encoding('gpt2')\n        tokens = enc.encode(text)\n        self.tokens = torch.tensor(tokens)\n        print(f\"Loaded {len(self.tokens)} tokens\")\n        print(f\"1 epoch = {len(self.tokens) \/\/ (B * T)} batches\")\n\n        # making changes in below code to accomodate the DDP and MultiGPU training\n        # data splitting\n        self.current_position = self.B * self.T * self.process_rank # for each process it's batch will start at rank times B times T\n\n    def next_batch(self):\n        # as well as makinng the changes in below code to always load the data on corresponding GPU accordingly \n        # and current position is advanced in such a way that it get's diffent data from every other GPU always\n        B, T = self.B, self.T\n        buf = self.tokens[self.current_position : self.current_position + B * T + 1]\n        # buf.to(dtype = torch.float16)\n        x = (buf[:-1]).view(B,T) # inputs\n        y = (buf[1:]).view(B,T) # targets\n        # advance the position in the tensor\n        self.current_position += B * T * self.num_processes\n        # if loading the next batch would be out of bounds, reset\n        if self.current_position + (B * T * self.num_processes + 1) &gt; len(self.tokens):\n            self.current_position = self.B * self.T * self.process_rank\n        return x,y\n<\/code><\/pre>\n<p>\u5f53\u4f7f\u7528\u68af\u5ea6\u7d2f\u79ef\u548c\u5fae\u6b65\u6765\u52a0\u901f\u6bcf\u4e2a\u65f6\u671f\u7684\u8bad\u7ec3\u65f6\uff0c\u4f60\u53ef\u80fd\u4e0d\u5e0c\u671b\u5728\u6bcf\u4e2a\u5fae\u6b65\u4e4b\u540e\u540c\u6b65\u68af\u5ea6\u3002\u76f8\u53cd\uff0c\u4f60\u66f4\u613f\u610f\u5728\u5b8c\u6210\u6574\u4e2a\u7d2f\u79ef\u6b65\u9aa4\u540e\u624d\u5728\u6240\u6709\u8fdb\u7a0b\u4e2d\u540c\u6b65\u5b83\u4eec\u3002\u4f46\u662f\uff0c\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0cPyTorch DDP \u4f1a\u5728\u6bcf\u6b21 loss.backward() \u8c03\u7528\u671f\u95f4\u540c\u6b65\u68af\u5ea6\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cPyTorch DDP \u63d0\u4f9b\u4e86 no_sync() \u4e0a\u4e0b\u6587\u7ba1\u7406\u5668\u3002\u6216\u8005\uff0c\u4f60\u53ef\u4ee5\u901a\u8fc7\u76f4\u63a5\u4fee\u6539 require_backward_grad_sync \u53d8\u91cf\u6765\u5b9e\u73b0\u76f8\u540c\u7684\u884c\u4e3a\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n<h2>6\u3001\u7ed3\u675f\u8bed<\/h2>\n<p>\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u662f\u8bad\u7ec3\u5e9e\u5927\u6a21\u578b\u7684\u5f3a\u5927\u5de5\u5177\uff0c\u5c24\u5176\u662f\u5f53\u4f60\u53ef\u4ee5\u8bbf\u95ee\u591a\u4e2a GPU \u65f6\u3002\u5b83\u7b80\u5316\u4e86\u6d41\u7a0b\u5e76\u6709\u6548\u5730\u5728\u8bbe\u5907\u4e4b\u95f4\u5206\u914d\u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n<p>\u9664\u4e86 DDP\uff0c\u8fd8\u6709\u5176\u4ed6\u53ef\u7528\u4e8e\u8bad\u7ec3\u5927\u578b\u6a21\u578b\u7684\u9ad8\u7ea7\u6280\u672f\uff0c\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u5b8c\u5168\u5206\u7247\u6570\u636e\u5e76\u884c\u8bad\u7ec3 (FSDP)\uff1a\u4e00\u79cd\u5728\u8bbe\u5907\u4e4b\u95f4\u5206\u7247\u6a21\u578b\u6743\u91cd\u548c\u4f18\u5316\u5668\u72b6\u6001\u7684\u6280\u672f\uff0c\u53ef\u5b9e\u73b0\u9ad8\u6548\u7684\u5185\u5b58\u4f7f\u7528\u3002<\/li>\n<li>\u5f20\u91cf\u5e76\u884c (TP)\uff1a\u5c06\u6a21\u578b\u7684\u5404\u4e2a\u5c42\u62c6\u5206\u5230\u591a\u4e2a\u8bbe\u5907\uff0c\u5141\u8bb8\u5728\u5c42\u5185\u8fdb\u884c\u5e76\u884c\u8ba1\u7b97\u3002<\/li>\n<li>\u7ba1\u9053\u5e76\u884c (PP)\uff1a\u5c06\u6a21\u578b\u5212\u5206\u4e3a\u8fde\u7eed\u9636\u6bb5\uff0c\u6bcf\u4e2a\u9636\u6bb5\u5206\u914d\u7ed9\u4e0d\u540c\u7684\u8bbe\u5907\uff0c\u4ece\u800c\u4fc3\u8fdb\u6a21\u578b\u5e76\u884c\u3002<\/li>\n<\/ul>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u6700\u8fd1\uff0c\u5728\u4f7f\u7528 OpenAI \u7684\u8bba\u6587\u201c\u8bed\u8a00\u6a21\u578b\u662f\u65e0\u76d1\u7763\u7684\u591a\u4efb\u52a1\u5b66\u4e60\u8005\u201d\u548c Andrej Karpathy \u7684 YouTube \u89c6\u9891\u201c\u8ba9\u6211\u4eec\u91cd\u73b0 GPT-2 (124M)\u201d\u4ece\u5934\u91cd\u73b0 GPT-2 LLM \u65f6\uff0c\u6211\u5f3a\u70c8\u5730\u60f3\u8981\u4e86\u89e3\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u7684\u5de5\u4f5c\u539f\u7406\u3002\u8bad\u7ec3\u5982\u6b64\u5927\u7684\u6a21\u578b\u9700\u8981\u591a GPU \u8bbe\u7f6e\uff0c\u800c\u4e14\u7531\u4e8e\u8fd9\u662f\u6211\u7b2c\u4e00\u6b21\u5c1d\u8bd5\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u8fd9\u79cd\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u6240\u4ee5\u8fd9\u4e2a\u4e3b\u9898\u5bf9\u6211\u6765\u8bf4\u662f\u5168\u65b0\u7684\u3002 \u4e3a\u4e86\u5f25\u8865\u8fd9\u4e00\u77e5\u8bc6\u5dee\u8ddd\uff0c\u6211\u7acb\u5373\u9605\u8bfb\u4e86 PyTorch \u7684 DDP \u6587\u6863\u5e76\u7cfb\u7edf\u5730\u7406\u89e3\u5b83\u3002\u672c\u6587\u5c31\u662f\u8fd9\u6bb5\u5b66\u4e60\u4e4b\u65c5\u7684\u6210\u679c\u3002 \u968f\u7740\u6570\u636e\u96c6\u548c\u6a21\u578b\u53d8\u5f97\u8d8a\u6765\u8d8a\u5927\uff0c\u5728\u591a\u4e2a GPU \u4e0a\u5206\u914d\u5de5\u4f5c\u8d1f\u8f7d\u4e0d\u4ec5\u6709\u7528\uff0c\u800c\u4e14\u5fc5\u4e0d\u53ef\u5c11\u3002\u5b83\u663e\u8457\u51cf\u5c11\u4e86\u8bad\u7ec3\u65f6\u95f4\uff0c\u589e\u5f3a\u4e86\u53ef\u6269\u5c55\u6027\uff0c\u5e76\u4f7f\u8bad\u7ec3\u5927\u89c4\u6a21\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002PyTorch \u7684\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c (DDP) \u662f\u6ee1\u8db3\u8fd9\u4e9b\u9700\u6c42\u7684\u5f3a\u5927\u89e3\u51b3\u65b9\u6848\u4e4b\u4e00\u3002 \u5728\u672c\u6587\u4e2d\uff0c\u6211\u5c06\u89e3\u91ca\u6211\u5bf9 DDP \u7684\u7406\u89e3\u3001\u5b83\u76f8\u5bf9\u4e8e\u6570\u636e\u5e76\u884c (DP) \u7684\u4f18\u52bf\u3001\u5b83\u7684\u5185\u90e8\u5de5\u4f5c\u539f\u7406\u4ee5\u53ca\u4f7f\u7528 ToyModel \u7684\u5b9e\u9645\u5b9e\u73b0\u793a\u4f8b\u3002 \u4f7f\u7528 Dall-E \u521b\u5efa\u7684\u6a21\u578b\u5e76\u884c\u6027\u56fe\u50cf \u662f PyTorch \u4e2d\u7684\u4e00\u4e2a\u5f3a\u5927\u6a21\u5757\uff0c\u5b83\u5141\u8bb8\u6211\u4eec\u5728\u591a\u53f0\u673a\u5668\u4e0a\u5e76\u884c\u5316\u6211\u4eec\u7684\u6a21\u578b\uff0c\u901a\u8fc7\u5728\u591a\u4e2a GPU \u4e0a\u590d\u5236\u6a21\u578b\u6765\u5b9e\u73b0\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002\u6bcf\u4e2a\u8fdb\u7a0b\u5728\u6570\u636e\u5b50\u96c6\u4e0a\u8bad\u7ec3\u6a21\u578b\u7684\u526f\u672c\uff0c\u5e76\u4e14\u68af\u5ea6\u5728\u8fdb\u7a0b\u4e4b\u95f4\u540c\u6b65\u4ee5\u786e\u4fdd\u6a21\u578b\u53c2\u6570\u7684\u4e00\u81f4\u66f4\u65b0\u3002 \u5728\u8bad\u7ec3\u5927\u578b\u6a21\u578b\u65f6\uff0c\u5355\u4e2a GPU \u901a\u5e38\u7f3a\u4e4f\u6709\u6548\u5904\u7406\u4efb\u52a1\u6240\u9700\u7684\u5185\u5b58\u6216\u8ba1\u7b97\u80fd\u529b\u3002\u5373\u4f7f\u53ef\u4ee5\u5728\u5355\u4e2a GPU \u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6240\u9700\u7684\u65f6\u95f4\u4e5f\u53ef\u80fd\u975e\u5e38\u957f\u3002\u4f7f\u7528\u591a\u4e2a GPU \u53ef\u4ee5\u5e2e\u52a9\u6211\u4eec\uff1a \u5904\u7406\u66f4\u5927\u7684\u6a21\u578b\uff1a\u5728 GPU \u4e4b\u95f4\u805a\u5408\u5185\u5b58\u4ee5\u8bad\u7ec3\u65e0\u6cd5\u5bb9\u7eb3\u5728\u5355\u4e2a GPU \u5185\u5b58\u4e2d\u7684\u6a21\u578b\u3002 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53736","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53736","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53736"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53736\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53736"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53736"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53736"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}