{"id":34518,"date":"2024-12-08T04:56:24","date_gmt":"2024-12-07T20:56:24","guid":{"rendered":"https:\/\/17aitech.com\/?p=34518"},"modified":"2024-12-08T04:56:24","modified_gmt":"2024-12-07T20:56:24","slug":"%e4%ba%94%e5%b9%b4%e5%90%8e%e7%9a%84%e4%bb%8a%e5%a4%a9%ef%bc%8c%e8%ae%ad%e7%bb%83gpt-2%e5%8f%aa%e9%9c%80%e4%b8%8d%e5%88%b0700%e5%88%80%e3%80%8124%e5%b0%8f%e6%97%b6%ef%bc%8ckarpathy%e5%8f%88%e6%95%b4","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34518","title":{"rendered":"\u4e94\u5e74\u540e\u7684\u4eca\u5929\uff0c\u8bad\u7ec3GPT-2\u53ea\u9700\u4e0d\u5230700\u5200\u300124\u5c0f\u65f6\uff0cKarpathy\u53c8\u6574\u65b0\u6d3b"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4e94\u5e74\u540e\u7684\u4eca\u5929\uff0c\u8bad\u7ec3GPT-2\u53ea\u9700\u4e0d\u5230700\u5200\u300124\u5c0f\u65f6\uff0cKarpathy\u53c8\u6574\u65b0\u6d3b<\/a><\/p>\n

\n
\n
\n

\u8bba\u8001\u9ec4\u5356\u94f2\u5b50\u7684\u6280\u672f\u542b\u91cf\u3002<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n

2019 \u5e74 2 \u6708\uff0cOpenAI \u53d1\u5e03\u4e86 GPT-2<\/mark><\/a>\uff0c\u56e0\u4e3a\u5728\u6587\u672c\u751f\u6210<\/mark>\u4e0a\u7684\u4f18\u5f02\u8868\u73b0\uff0c\u4ee5\u53ca\u5bf9\u4e8e\u9884\u8bad\u7ec3 Transformer \u67b6\u6784\u7684\u5145\u5206\u8fd0\u7528\uff0c\u88ab\u8ba4\u4e3a\u662f\u5982\u4eca\u5927\u9884\u8a00\u6a21\u578b\u7684\u300c\u59cb\u7956\u300d\u3002<\/p>\n

\u4e94\u5e74\u540e\u7684\u4eca\u5929\uff0c\u8bad\u7ec3 GPT-2<\/mark> \u8fd9\u6837 15 \u4ebf\u53c2\u6570<\/mark>\u7684\u5927\u6a21\u578b\uff0c\u53ea\u9700\u8981\u82b1\u8d39 672 \u7f8e\u5143\uff0c\u5728\u4e00\u4e2a 8XH100 \u7684 GPU \u8282\u70b9\u4e0a\u8dd1 24 \u4e2a\u5c0f\u65f6\u5c31\u53ef\u4ee5\u641e\u5b9a\u4e86\u3002<\/p>\n

\u672c\u5468\u56db\uff0c\u524d\u7279\u65af\u62c9 Autopilot \u8d1f\u8d23\u4eba\u3001OpenAI \u79d1\u5b66\u5bb6 Andrej Karpathy \u5728\u4ed6\u7eaf C \u8bed\u8a00\u590d\u73b0 GPT-2<\/mark> \u5927\u6a21\u578b\u7684\u9879\u76ee\u300cllm.c\u300d<\/a>\u7684\u6700\u65b0\u8fdb\u5c55\u4e2d\u5206\u4eab\u4e86\u4ed6\u7684\u8bad\u7ec3\u5fc3\u5f97\uff1a<\/p>\n

<\/a><\/section>\n

\u4ee4\u4eba\u96be\u4ee5\u7f6e\u4fe1\u7684\u662f\uff0c\u7531\u4e8e\u8ba1\u7b97\u786c\u4ef6\uff08\u82f1\u4f1f\u8fbe H100 GPU\uff09\u3001\u8f6f\u4ef6\uff08CUDA\u3001cuBLAS\u3001cuDNN\u3001FlashAttention \u7b49\uff09\u548c\u6570\u636e\u8d28\u91cf\uff08\u4f8b\u5982 FineWeb-Edu \u6570\u636e\u96c6\uff09\u7684\u6539\u8fdb\uff0c\u8fc7\u53bb 5 \u5e74\u95f4\uff0c\u5927\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u8bad\u7ec3\u6210\u672c\u5927\u5e45\u4e0b\u964d\u3002Karpathy \u8868\u793a\uff0c\u5bf9\u4e8e\u6b64\u6b21\u5b9e\u8df5\uff0c\u7b97\u6cd5\u9075\u5faa GPT-2<\/mark>\/3 \u8bba\u6587\u57fa\u672c\u4fdd\u6301\u539f\u6837\u4e0d\u53d8\u3002<\/p>\n

\u5f53\u5e74 OpenAI \u8bad\u7ec3 GPT-2<\/mark> \u82b1\u8d39\u4e86\u591a\u5c11\u94b1\uff1f\u8fd9\u662f\u4e2a\u81f3\u4eca\u4ecd\u7136\u672a\u77e5\u7684\u6570\u5b57\u3002Karpathy \u7c97\u7565\u5730\u4f30\u7b97\u8ba4\u4e3a\u662f\u8fd9\u56de\u6210\u672c\u7684 100 \u500d\uff0c\u5927\u6982\u8981\u5230 10 \u4e07\u7f8e\u5143\u7684\u91cf\u7ea7\u3002<\/p>\n

<\/a><\/section>\n

\u57fa\u672c\u76f8\u540c\u7684\u4efb\u52a1\uff0c\u8fd0\u884c\u6548\u7387\u5374\u6709\u5929\u58e4<\/mark>\u4e4b\u522b\uff0c\u8fd9\u4f53\u73b0\u4e86\u8fd1\u51e0\u5e74\u6765 AI \u9886\u57df\u548c\u7b97\u529b\u57fa\u7840\u8bbe\u65bd\u7684\u98de\u901f\u53d1\u5c55\u3002<\/p>\n

\u7531\u4e8e llm.c \u662f\u5728 C\/CUDA \u4e2d GPT \u8bad\u7ec3\u7684\u76f4\u63a5\u5b9e\u73b0\uff0c\u56e0\u6b64\u8981\u6c42\u5176\u5b9e\u5f88\u5c11 \u2014\u2014 \u4e0d\u9700\u8981 conda \u73af\u5883\u3001Python \u89e3\u91ca\u5668\u3001pip \u5b89\u88c5\u7b49\u3002\u5982\u679c\u4f60\u4e5f\u8981\u5c1d\u8bd5\uff0c\u53ef\u4ee5\u542f\u52a8\u4e91 GPU \u8282\u70b9\uff08\u4f8b\u5982\u5728 Lambda \u4e0a\uff09\uff0c\u53ef\u9009\u62e9\u5b89\u88c5 NVIDIA cuDNN\u3001NCCL\/MPI\uff0c\u4e0b\u8f7d .bin \u6570\u636e\u5206\u7247\uff0c\u7f16\u8bd1\u5e76\u8fd0\u884c\uff0c\u51e0\u5206\u949f\u540e\u5373\u53ef\u5f00\u59cb\u3002<\/p>\n

\u7136\u540e\uff0c\u4f60\u5c31\u53ef\u4ee5\u7b49\u5f85 24 \u5c0f\u65f6\uff0c\u7136\u540e\u6b23\u8d4f\u901a\u7528\u5927\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u80fd\u529b\u4e86\u3002<\/p>\n

\u300c\u5bf9\u4e8e llm.c \u9879\u76ee\u6765\u8bf4\uff0c\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u597d\u7684\u8282\u70b9\u3002\u56e0\u4e3a\u6574\u4e2a\u9879\u76ee\u90fd\u662f\u4ece\u6211\u8003\u8651\u4e3a\u6559\u80b2\u89c6\u9891\u91cd\u73b0 GPT-2<\/mark> \u5f00\u59cb\u7684\u3002\u6211\u9047\u5230\u4e00\u4e9b PyTorch \u7684\u4e1c\u897f\u65f6\u5361\u4f4f\u4e86\uff0c\u7136\u540e\u6124\u6012\u5730\u9000\u51fa\uff0c\u518d\u7528 C\/CUDA \u4ece\u5934\u5f00\u59cb\u7f16\u5199\u6574\u4e2a\u9879\u76ee\uff0c\u300dKarpathy \u8868\u793a\u3002\u300c\u8fd9\u8ba9\u6211\u8e0f\u4e0a\u4e86\u6bd4\u9884\u60f3\u66f4\u957f\u7684\u65c5\u7a0b\u3002\u4f46\u5b83\u975e\u5e38\u6709\u8da3\uff0c\u6211\u5b66\u5230\u4e86\u66f4\u591a\u7684 CUDA\uff0c\u4e00\u8def\u4e0a\u7ed3\u4ea4\u4e86\u670b\u53cb\uff0c\u73b0\u5728\u7684 llm.c \u771f\u7684\u5f88\u68d2\u3002\u5b83\u6709\u5927\u7ea6 5000 \u884c\u4ee3\u7801\uff0c\u7f16\u8bd1\u548c\u6b65\u9aa4\u975e\u5e38\u5feb\uff0c\u51e0\u4e4e\u4e0d\u9700\u8981\u7b49\u5f85\u3002\u5b83\u5177\u6709\u6052\u5b9a\u7684\u5185\u5b58\u5360\u7528\uff0c\u5b83\u4ee5\u6df7\u5408\u7cbe\u5ea6\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u7528 NNCL \u5206\u5e03\u5728\u591a\u8282\u70b9\u4e0a\u3002\u5b83\u662f\u6309\u4f4d\u786e\u5b9a\u6027\u7684\uff0c\u5e76\u4e14\u5f98\u5f8a\u5728 MFU \u7684 50% \u5de6\u53f3\u3002\u6240\u4ee5\u5b83\u5f88 ok\u3002\u300d<\/p>\n

\u5bf9\u4e8e llm.c \u9879\u76ee\u800c\u8a00\uff0c\u8d8a\u505a\u4f3c\u4e4e\u6316\u5f97\u5751\u8d8a\u5927\u3002Andrej Karpathy \u5bf9\u76ee\u524d\u7684\u8fd0\u884c\u7ed3\u679c\u4ecd\u7136\u4e0d\u662f 100% \u6ee1\u610f \u2014\u2014 \u4ed6\u8ba4\u4e3a\u8bc4\u4f30\u5e94\u8be5\u66f4\u597d\uff0c\u8bad\u7ec3\u5e94\u8be5\u66f4\u7a33\u5b9a\uff0c\u5c24\u5176\u662f\u5728\u8f83\u957f\u65f6\u95f4\u8fd0\u884c\u7684\u8f83\u5927\u6a21\u578b\u5c3a\u5bf8\u4e0b\u3002<\/p>\n

\u4ed6\u8fd8\u9884\u544a\u4e86\u4e00\u4e9b\u6709\u8da3\u7684\u65b0\u65b9\u5411\uff1afp8\uff08\u5373\u5c06\u63a8\u51fa\uff09\u3001\u63a8\u7406\u3001\u5fae\u8c03\u3001\u591a\u6a21\u6001\uff08VQVAE \u7b49\uff09\u3001\u66f4\u73b0\u4ee3\u7684\u67b6\u6784\uff08Llama\/Gemma\uff09\u3002llm.c \u7684\u76ee\u6807\u4ecd\u7136\u662f\u4e3a\u529f\u80fd\u9f50\u5168\u7684 LLM \u667a\u80fd\u4f53\u63d0\u4f9b\u7b80\u5355\u3001\u6700\u5c0f\u3001\u5e72\u51c0\u7684\u8bad\u7ec3\u5806\u6808\uff0c\u76f4\u63a5\u4f7f\u7528 C\/CUDA\uff0c\u5e76\u5305\u542b\u914d\u5957\u7684\u6559\u80b2\u6750\u6599\uff0c\u53ef\u4ee5\u8ba9\u8bb8\u591a\u521d\u5b66\u8005\u5feb\u901f\u4e86\u89e3\u8fd9\u4e2a\u4ee4\u4eba\u656c\u754f\u7684\u9886\u57df\u3002<\/p>\n

\u8bf4\u5b8c\u4e86\u8fd9\u4e48\u591a\uff0c\u8be5\u770b\u770b 24 \u5c0f\u65f6\u8bad\u7ec3 GPT-2<\/mark> \u7684\u6210\u679c\u4e86\uff1aKarpathy \u4f7f\u7528\u66f4\u957f\u7684 400B token GPT-2<\/mark> \u8fd0\u884c\uff08\u4ece 33B token \u589e\u52a0\uff09\uff0c\u6548\u679c\u826f\u597d\uff0c\u76f4\u5230 330B\uff08\u8fbe\u5230 61% HellaSwag\uff0c\u8fdc\u9ad8\u4e8e\u8fd9\u4e2a\u5927\u5c0f\u7684 GPT-2<\/mark> \u548c GPT-3\uff09\uff0c\u7136\u540e\u5728\u8fd9\u4e2a\u56fe\u4e4b\u540e\u4e0d\u4e45\u7206\u70b8\u4e86\u3002\u76ee\u524d\u4f5c\u8005\u8fd8\u5728\u7ee7\u7eed\u8fdb\u884c\u7814\u7a76\u3002<\/p>\n

<\/a><\/section>\n

\u63a5\u4e0b\u6765\u770b\u8be6\u7ec6\u9879\u76ee\u4ecb\u7ecd\u3002<\/p>\n

<\/a><\/section>\n

GitHub \u5730\u5740\uff1ahttps:\/\/github.com\/karpathy\/llm.c\/discussions\/677<\/p>\n

\u8bad\u7ec3<\/strong>\u3002\u4f7f\u7528 llm.c \u8bad\u7ec3 GPT-2<\/mark> \u975e\u5e38\u7b80\u5355\uff0c\u56e0\u4e3a\u5b83\u662f\u7528 C\/CUDA \u7f16\u5199\u7684\uff0c\u56e0\u6b64\u4e0d\u9700\u8981 minconda\u3001Python\u3001PyTorch \u7b49\u3002\u4f60\u53ea\u9700\u4e00\u4e2a 8XH100 GPU box\uff0cKarpathy \u5efa\u8bae\u4ece Lambda Labs \u8d2d\u4e70\u4e00\u4e2a\u3002<\/p>\n

\u4e0d\u8fc7 llm.c \u5728\u8ba1\u7b97\u4e0a\u5f88\u7075\u6d3b\uff0c\u5982\u679c\u4f60\u53ea\u6709 1 \u4e2a GPU\uff0c\u4ecd\u7136\u53ef\u4ee5\u8bad\u5f97 GPT-2<\/mark>\uff0c\u8fd9\u65f6\u4f60\u9700\u8981\u7b49\u5f85 8 \u5929\u800c\u4e0d\u662f 1 \u5929\u3002\u5982\u679c\u4f60\u6709 16 \u4e2a GPU\uff08\u4f8b\u5982\u4f7f\u7528\u65b0\u7684 Lambda 1 Click Clusters\uff09\uff0c\u5219\u80fd\u591f\u8bad\u7ec3\u591a\u8282\u70b9\uff0c\u8fd9\u65f6\u53ea\u9700\u7b49\u5f85 12 \u5c0f\u65f6\u3002\u542f\u52a8\u8282\u70b9\u540e\uff0c\u4ee5\u4e0b\u662f\u8bad\u7ec3 GPT-2<\/mark> \u7684\u5b8c\u6574\u8bf4\u660e\uff1a<\/p>\n

#\u00a0install\u00a0cudnn\u00a0so\u00a0we\u00a0can\u00a0use\u00a0FlashAttention\u00a0and\u00a0run\u00a0fast\u00a0(optional)\r\n#\u00a0https:\/\/developer.nvidia.com\/cudnn-downloads\r\n#\u00a0for\u00a0me,\u00a0CUDA\u00a012\u00a0(run\u00a0`nvcc\u00a0--version`)\u00a0running\u00a0on\u00a0Linux\u00a0x86_64\u00a0Ubuntu\u00a022.04\r\nwget\u00a0https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu2204\/x86_64\/cuda-keyring_1.1-1_all.deb\r\nsudo\u00a0dpkg\u00a0-i\u00a0cuda-keyring_1.1-1_all.deb\r\nsudo\u00a0apt-get\u00a0update\r\nsudo\u00a0apt-get\u00a0-y\u00a0install\u00a0libcudnn9-dev-cuda-12\r\n#\u00a0\"install\"\u00a0cudnn-frontend\u00a0to\u00a0~\/\r\ngit\u00a0clone\u00a0https:\/\/github.com\/NVIDIA\/cudnn-frontend.git\r\n#\u00a0install\u00a0MPI\u00a0(optional,\u00a0if\u00a0you\u00a0intend\u00a0to\u00a0use\u00a0multiple\u00a0GPUs)\r\n#\u00a0(you\u00a0might\u00a0also\u00a0have\u00a0to\u00a0install\u00a0NVIDIA\u00a0NCCL\u00a0if\u00a0it\u00a0doesn't\u00a0come\u00a0with\u00a0your\u00a0setup)\r\nsudo\u00a0apt\u00a0-y\u00a0install\u00a0openmpi-bin\u00a0openmpi-doc\u00a0libopenmpi-dev\r\n#\u00a0download\u00a0and\u00a0enter\u00a0llm.c\u00a0repo\r\ngit\u00a0clone\u00a0https:\/\/github.com\/karpathy\/llm.c.gitcd\u00a0llm.c\r\n#\u00a0download\u00a0the\u00a0\"starter\u00a0pack\"\u00a0(~1GB\u00a0download)\r\n#\u00a0contains\u00a0GPT2-124M\u00a0weights\u00a0(used\u00a0in\u00a0tests),\u00a0tokenizer,\u00a0eval\u00a0data\u00a0.bin\u00a0s\r\n.\/dev\/download_starter_pack.sh\r\n#\u00a0download\u00a0the\u00a0training\u00a0dataset\u00a0(FineWeb-Edu\u00a0100B\u00a0token)\u00a0.bin\u00a0data\u00a0shards\r\n#\u00a0note:\u00a0this\u00a0is\u00a0a\u00a0total\u00a0of\u00a01001\u00a0data\u00a0shards.\u00a0If\u00a0you\u00a0only\u00a0want\u00a0to\u00a0test\u00a0things\r\n#\u00a0out\u00a0and\u00a0don't\u00a0want\u00a0to\u00a0do\u00a0an\u00a0actual\u00a0run,\u00a0feel\u00a0free\u00a0to\u00a0append\u00a0the\u00a0number\u00a0of\r\n#\u00a0training\u00a0shards\u00a0to\u00a0download\u00a0(e.g.\u00a0for\u00a0just\u00a010\u00a0shards:\u00a0.\/edu_fineweb.sh\u00a010)\r\n#\u00a0the\u00a0full\u00a0dataset\u00a0is\u00a0~200GB,\u00a0we\u00a0can\u00a0store\u00a0it\u00a0here\u00a0in\u00a0dev\/data\u00a0directory.\r\ncd\u00a0dev\/data\r\n.\/edu_fineweb.sh\r\n#\u00a0compile\u00a0(~1\u00a0min\u00a01st\u00a0time\u00a0for\u00a0cuDNN\u00a0mostly,\u00a0few\u00a0sec\u00a0from\u00a0then\u00a0on)\r\ncd\u00a0..\/..\/\r\nmake\u00a0train_gpt2cu\u00a0USE_CUDNN=1\r\n#\u00a0and\u00a0train!\u00a0(wait\u00a024\u00a0hours\u00a0here)\r\nmpirun\u00a0-np\u00a08\u00a0.\/train_gpt2cu\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-i\u00a0\"dev\/data\/edu_fineweb100B\/edu_fineweb_train_*.bin\"\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-j\u00a0\"dev\/data\/edu_fineweb100B\/edu_fineweb_val_*.bin\"\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-o\u00a0\"log_gpt2_1558M\"\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-v\u00a0250\u00a0-s\u00a0300000\u00a0-g\u00a0384\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-h\u00a01\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-b\u00a016\u00a0-t\u00a01024\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-d\u00a01048576\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-r\u00a00\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-z\u00a01\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-c\u00a00.1\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-k\u00a0\"cosine\"\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-l\u00a00.0006\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-q\u00a00.1\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-u\u00a0700\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-n\u00a02000\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-x\u00a032000\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-ge\u00a01\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-y\u00a01\u00a0\r\n\u00a0\u00a0\u00a0\u00a0-e\u00a0\"d48\"\r\n<\/code><\/pre>\n
\u5f00\u59cb\u4f18\u5316\uff1a<\/p>\n
num_parameters:\u00a01557686400\u00a0=>\u00a0bytes:\u00a03115372800\r\nallocated\u00a02971\u00a0MiB\u00a0for\u00a0model\u00a0parameters\r\nbatch_size\u00a0B=16\u00a0*\u00a0seq_len\u00a0T=1024\u00a0*\u00a0num_processes=8\u00a0and\u00a0total_batch_size=1048576\r\n=>\u00a0setting\u00a0grad_accum_steps=8\r\ncreated\u00a0directory:\u00a0log_gpt2_1558M\r\nallocating\u00a040409\u00a0MiB\u00a0for\u00a0activations\r\nval\u00a0loss\u00a011.129390\r\nallocating\u00a02971\u00a0MiB\u00a0for\u00a0parameter\u00a0gradients\r\nallocating\u00a0742\u00a0MiB\u00a0for\u00a0AdamW\u00a0optimizer\u00a0state\u00a0m\r\nallocating\u00a0742\u00a0MiB\u00a0for\u00a0AdamW\u00a0optimizer\u00a0state\u00a0v\r\nallocating\u00a0742\u00a0MiB\u00a0for\u00a0master\u00a0copy\u00a0of\u00a0params\r\nstep\u00a0\u00a0\u00a0\u00a01\/32000\u00a0|\u00a0loss\u00a011.133732\u00a0(+nanz)|\u00a0norm\u00a052.9732\u00a0(+nanz)|\u00a0lr\u00a08.57e-07\u00a0|\u00a03056.36\u00a0ms\u00a0|\u00a042.6%\u00a0bf16\u00a0MFU\u00a0|\u00a0343080\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a02\/32000\u00a0|\u00a0loss\u00a010.539388\u00a0(+nanz)|\u00a0norm\u00a043.5996\u00a0(+nanz)|\u00a0lr\u00a01.71e-06\u00a0|\u00a02747.19\u00a0ms\u00a0|\u00a047.4%\u00a0bf16\u00a0MFU\u00a0|\u00a0381690\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a03\/32000\u00a0|\u00a0loss\u00a09.894109\u00a0(+nanz)|\u00a0norm\u00a023.2229\u00a0(+nanz)|\u00a0lr\u00a02.57e-06\u00a0|\u00a02753.25\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381259\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a04\/32000\u00a0|\u00a0loss\u00a09.566241\u00a0(+nanz)|\u00a0norm\u00a028.4920\u00a0(+nanz)|\u00a0lr\u00a03.43e-06\u00a0|\u00a02741.47\u00a0ms\u00a0|\u00a047.5%\u00a0bf16\u00a0MFU\u00a0|\u00a0381690\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a05\/32000\u00a0|\u00a0loss\u00a09.482848\u00a0(+nanz)|\u00a0norm\u00a023.7817\u00a0(+nanz)|\u00a0lr\u00a04.29e-06\u00a0|\u00a02752.07\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381507\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a06\/32000\u00a0|\u00a0loss\u00a09.332832\u00a0(+nanz)|\u00a0norm\u00a015.9113\u00a0(+nanz)|\u00a0lr\u00a05.14e-06\u00a0|\u00a02751.01\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381431\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a07\/32000\u00a0|\u00a0loss\u00a09.165650\u00a0(+nanz)|\u00a0norm\u00a010.5941\u00a0(+nanz)|\u00a0lr\u00a06.00e-06\u00a0|\u00a02753.03\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381327\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a08\/32000\u00a0|\u00a0loss\u00a09.132234\u00a0(+nanz)|\u00a0norm\u00a016.2733\u00a0(+nanz)|\u00a0lr\u00a06.86e-06\u00a0|\u00a02748.91\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381348\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a0\u00a09\/32000\u00a0|\u00a0loss\u00a09.097384\u00a0(+nanz)|\u00a0norm\u00a012.1342\u00a0(+nanz)|\u00a0lr\u00a07.71e-06\u00a0|\u00a02748.73\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381367\u00a0tok\/s\r\nstep\u00a0\u00a0\u00a010\/32000\u00a0|\u00a0loss\u00a09.072879\u00a0(+nanz)|\u00a0norm\u00a010.5923\u00a0(+nanz)|\u00a0lr\u00a08.57e-06\u00a0|\u00a02749.40\u00a0ms\u00a0|\u00a047.3%\u00a0bf16\u00a0MFU\u00a0|\u00a0381369\u00a0tok\/s\r\n...\r\n<\/code><\/pre>\n
\u6bcf\u4e00\u6b65\u5927\u7ea6\u9700\u8981 2.75 \u79d2\uff0c\u5171\u6709 32000 \u6b65\uff0c\u6240\u4ee5\u73b0\u5728\u6211\u4eec\u7b49\u5f85 24 \u5c0f\u65f6\u5de6\u53f3\u3002\u5728\u6bcf\u4e00\u6b65\u4e2d\uff0c\u8bad\u7ec3\u8fd0\u884c\u90fd\u4f1a\u5360\u7528\u7ea6 100 \u4e07\u4e2a FineWeb-EDU token\uff08\u8fd9\u4e9b\u6765\u81ea\u4e92\u8054\u7f51\u7684\u6559\u80b2\u7f51\u9875\uff09\uff0c\u5e76\u66f4\u65b0\u6a21\u578b\u7684 15.58 \u4ebf\u4e2a\u6743\u91cd<\/mark>\uff0c\u4f7f\u5176\u80fd\u591f\u66f4\u597d\u5730\u9884\u6d4b\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a token\u3002\u6700\u540e\u5c06\u603b\u5171\u5904\u7406 32000 * 1048576 = 33.6B \u4e2a token\u3002<\/strong>\u968f\u7740\u66f4\u597d\u5730\u9884\u6d4b\u4e0b\u4e00\u4e2a token\uff0c\u635f\u5931\u4f1a\u4e0b\u964d\u3002\u8303\u6570<\/mark>\u5c06\u7a33\u5b9a\u5728 0.1-1 \u5de6\u53f3\uff0c\u5b66\u4e60\u7387<\/mark>\u5728\u524d\u9762\u51e0\u6b65\u9884\u70ed\u3002\u6a21\u578b flops \u5229\u7528\u7387 (MFU) \u7ea6\u4e3a 50%\uff0c\u975e\u5e38\u9ad8\u6548\u3002<\/p>\n
\u7b49\u5f85 24 \u5c0f\u65f6\u540e\uff0c\u5c31\u53ef\u4ee5\u4f7f\u7528 dev\/vislog.ipynb jupyter \u7b14\u8bb0\u672c\u53ef\u89c6\u5316 main.log \u65e5\u5fd7\u6587\u4ef6\u3002\u4e3a\u6b64\uff0c\u4f60\u8fd8\u9700\u8981\u5b89\u88c5 Python \u548c matplotlib\u3002<\/p>\n
\u53c2\u6570<\/mark>\u6307\u5357<\/strong>\u3002OpenAI \u53d1\u5e03\u7684 GPT-2<\/mark> \u5305\u542b\u6a21\u578b\u6743\u91cd<\/mark>\uff0c\u4f46\u7ec6\u8282\u5f88\u5c11\uff1b\u800c GPT-3 \u7248\u672c\u6ca1\u6709\u6743\u91cd<\/mark>\uff0c\u4f46\u7ec6\u8282\u5f88\u591a\u3002\u56e0\u6b64\uff0c\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u9075\u5faa GPT-3 \u8bba\u6587\u8d85\u53c2\u6570<\/mark>\uff0c\u56e0\u4e3a GPT-2<\/mark> \u8bba\u6587\u7684\u4fe1\u606f\u975e\u5e38\u5c11\u3002\u5177\u4f53\u53c2\u89c1\u539f\u9879\u76ee\u3002<\/p>\n
\u5185\u5b58\u6307\u5357<\/strong>\u3002\u5927\u591a\u6570\u4eba\u53ef\u80fd\u9762\u4e34\u7684\u6700\u5927\u9650\u5236\u662f\u4ed6\u4eec\u7684 GPU \u6ca1\u6709 80GB\u3002\u6ca1\u5173\u7cfb\uff0c\u4f60\u4ecd\u7136\u53ef\u4ee5\u8fd0\u884c\u4e0a\u9762\u7684\u6240\u6709\u5185\u5bb9\uff0c\u53ea\u662f\u8fd0\u884c\u901f\u5ea6\u4f1a\u66f4\u6162\u3002\u56e0\u6b64\u5982\u679c\u6a21\u578b\u4e0d\u9002\u914d\uff0c\u4f60\u4f1a\u600e\u4e48\u505a\uff1f\u6700\u91cd\u8981\u7684\u662f\u5fae\u6279\u5927\u5c0f – b\u3002\u5c1d\u8bd5\u51cf\u5c0f\u5b83\uff0c\u4f46\u5c06\u5176\u4fdd\u6301\u5728\u5408\u9002\u7684\u6570\u5b57\uff0c\u4f8b\u5982 16 \u2192 8 \u2192 4 \u2192 2 \u2192 1\u3002\u4ece\u90a3\u91cc\u5f00\u59cb\uff0c\u5c1d\u8bd5\u4f7f\u7528\u91cd\u8ba1\u7b97\u8bbe\u7f6e -r\uff0c\u5373 0\uff08\u6700\u5feb\u4e14\u6709\u5927\u91cf\u5185\u5b58\uff09\u30011\uff08\u7a0d\u5fae\u6162\u4e00\u70b9\uff0c\u4f46\u8282\u7701\u5927\u91cf\u5185\u5b58\uff09\u6216 2\uff08\u7a0d\u5fae\u6162\u4e00\u70b9\uff0c\u8282\u7701\u8f83\u5c11\u5185\u5b58\uff09\u3002<\/p>\n
\u4f60\u53ef\u4ee5\u505a\u7684\u4e0b\u4e00\u4ef6\u4e8b\u662f\u7981\u7528 fp32 \u4e2d\u7684\u4e3b\u6743\u91cd<\/mark>\uff0c\u53ef\u4ee5\u4f7f\u7528 – w 0 \uff08\u9ed8\u8ba4\u503c 1\uff09\u6765\u6267\u884c\u6b64\u64cd\u4f5c\u3002\u6211\u4eec\u4e0d\u4f1a\u7ef4\u62a4 fp32 \u53c2\u6570<\/mark>\u526f\u672c\u3002\u6839\u636e\u7ecf\u9a8c\uff0c\u5728\u4e4b\u524d\u7684\u51e0\u6b21\u8fd0\u884c\u4e2d\uff0c\u8fd9\u4f3c\u4e4e\u6ca1\u95ee\u9898\uff0c\u53ef\u80fd\u662f\u56e0\u4e3a\u4f7f\u7528\u4e86\u968f\u673a\u820d\u5165\u3002\u5982\u679c\u8fd8\u4e0d\u9002\u5408\uff0c\u5219\u53ef\u4ee5\u5c1d\u8bd5\u4f7f\u7528 -t \u00a0\u6765\u51cf\u5c11\u6700\u5927\u5e8f\u5217\u957f\u5ea6\uff0c\u9ed8\u8ba4\u503c\u4e3a 1024\uff0c\u4f60\u53ef\u4ee5\u5c06\u5176\u964d\u4f4e\u5230 512\u3001256 \u7b49\u3002\u4f46\u73b0\u5728\u4f60\u4f1a\u8ba9\u6a21\u578b\u53d8\u5f97\u66f4\u7cdf\uff0c\u56e0\u4e3a\u5b83\u7684\u6700\u5927\u6ce8\u610f\u529b\u8de8\u5ea6\u6b63\u5728\u51cf\u5c11\u3002<\/p>\n
\u4ee3\u7801<\/strong>\u3002Karpathy \u5bf9 llm.c \u7565\u6709\u504f\u7231\uff0c\u8ba4\u4e3a\u5b83\u975e\u5e38\u6f02\u4eae\uff1a<\/p>\n