{"id":4289,"date":"2024-07-12T13:47:34","date_gmt":"2024-07-12T05:47:34","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4289"},"modified":"2024-07-12T13:52:58","modified_gmt":"2024-07-12T05:52:58","slug":"opendiloco%ef%bc%9a%e7%94%a8%e4%ba%8e%e5%85%a8%e7%90%83%e5%88%86%e5%b8%83%e5%bc%8f%e4%bd%8e%e9%80%9a%e4%bf%a1%e8%ae%ad%e7%bb%83%e7%9a%84%e5%bc%80%e6%ba%90%e6%a1%86%e6%9e%b6","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4289","title":{"rendered":"OPENDILOCO\uff1a\u7528\u4e8e\u5168\u7403\u5206\u5e03\u5f0f\u4f4e\u901a\u4fe1\u8bad\u7ec3\u7684\u5f00\u6e90\u6846\u67b6"},"content":{"rendered":"\n<p>\u9686\u91cd\u63a8\u51fa OpenDiLoCo\uff0c\u8fd9\u662f DeepMind \u5206\u5e03\u5f0f\u4f4e\u901a\u4fe1 \uff08DiLoCo\uff09 \u65b9\u6cd5\u7684\u5f00\u6e90\u5b9e\u73b0\u548c\u6269\u5c55\uff0c\u53ef\u5b9e\u73b0\u5168\u7403\u5206\u5e03\u5f0f AI \u6a21\u578b\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u4e0a\u5468\uff0c\u6211\u4eec\u53d1\u5e03\u4e86<a href=\"https:\/\/www.primeintellect.ai\/blog\/introducing-prime-intellect\">\u603b\u4f53\u89c4\u5212<\/a>\u7684\u7b2c\u4e00\u6b65\uff0c\u542f\u52a8\u4e86<a href=\"https:\/\/www.primeintellect.ai\/blog\/compute\">Prime Intellect Compute Exchange<\/a>\uff0c\u4ee5\u805a\u5408\u548c\u7f16\u6392\u5168\u7403\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n\n\n\n<p>\u4eca\u5929\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5730\u5ba3\u5e03\uff0c\u901a\u8fc7\u5f00\u6e90\u6211\u4eec\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6\uff0c\u5728\u7b2c\u4e8c\u90e8\u5206\u5411\u524d\u8fc8\u51fa\u4e86\u4e00\u5927\u6b65\uff0c\u4ee5\u5b9e\u73b0\u8de8\u5168\u7403\u5206\u5e03\u5f0f\u786c\u4ef6\u7684\u534f\u4f5c\u6a21\u578b\u5f00\u53d1\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u63d0\u4f9b Deepmind \u7684 DiLoCo \u5b9e\u9a8c\u7684\u53ef\u91cd\u590d\u5b9e\u73b0\uff0c\u5e76\u5728\u53ef\u6269\u5c55\u7684\u53bb\u4e2d\u5fc3\u5316\u8bad\u7ec3\u6846\u67b6\u5185\u63d0\u4f9b\u3002\u6211\u4eec\u901a\u8fc7\u5728\u4e24\u5927\u6d32\u548c\u4e09\u4e2a\u56fd\u5bb6\/\u5730\u533a\u8bad\u7ec3\u6a21\u578b\u6765\u8bc1\u660e\u5176\u6709\u6548\u6027\uff0c\u540c\u65f6\u4fdd\u6301 90-95% \u7684\u8ba1\u7b97\u5229\u7528\u7387\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5c06 DiLoCo \u7684\u5927\u5c0f\u6269\u5c55\u5230\u539f\u59cb\u5de5\u4f5c\u7684 3 \u500d\uff0c\u8bc1\u660e\u4e86\u5b83\u5bf9\u5341\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u200d<\/p>\n\n\n\n<p>\u8bba\u6587\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2407.07852\">https:\/\/arxiv.org\/abs\/2407.07852<\/a><\/p>\n\n\n\n<p>\u4ee3\u7801\uff1a<a href=\"https:\/\/github.com\/PrimeIntellect-ai\/OpenDiLoCo\">https:\/\/github.com\/PrimeIntellect-ai\/OpenDiLoCo<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br>OpenDiLoCo \u8bad\u7ec3\u8fd0\u884c\u53ef\u89c6\u5316<\/h3>\n\n\n\n<figure class=\"wp-block-embed is-type-rich\"><\/figure>\n\n\n\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5df2\u7ecf\u5f7b\u5e95\u6539\u53d8\u4e86\u4eba\u5de5\u667a\u80fd\uff0c\u4f46\u4f20\u7edf\u4e0a\u8bad\u7ec3\u5b83\u4eec\u9700\u8981\u5927\u89c4\u6a21\u7684\u96c6\u4e2d\u5f0f\u8ba1\u7b97\u96c6\u7fa4\u3002\u8fd9\u79cd\u8d44\u6e90\u7684\u96c6\u4e2d\u9650\u5236\u4e86\u8c01\u53ef\u4ee5\u53c2\u4e0e\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\uff0c\u5e76\u51cf\u7f13\u4e86\u521b\u65b0\u7684\u6b65\u4f10\u3002<\/p>\n\n\n\n<p>\u6700\u8fd1\uff0c\u6211\u4eec\u53d1\u8868\u4e86\u4e00\u7bc7\u8be6\u7ec6\u7684\u535a\u5ba2\u6587\u7ae0\uff0c\u63a2\u8ba8\u4e86<a href=\"https:\/\/www.primeintellect.ai\/blog\/our-approach-to-decentralized-training\">\u53bb\u4e2d\u5fc3\u5316\u4eba\u5de5\u667a\u80fd\u8bad\u7ec3\u7684\u6700\u65b0\u6280\u672f<\/a>\u3002\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u91cd\u70b9\u4ecb\u7ecd\u4e86\u6700\u6709\u524d\u9014\u7684\u65b9\u6cd5\u548c\u4ecd\u9700\u8981\u514b\u670d\u7684\u51e0\u4e2a\u5173\u952e\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7f13\u6162\u7684\u4e92\u8fde\u5e26\u5bbd<\/li>\n\n\n\n<li>\u786e\u4fdd\u5bb9\u9519\u57f9\u8bad<\/li>\n\n\n\n<li>\u975e\u540c\u6784\u786c\u4ef6\u8bbe\u7f6e<\/li>\n\n\n\n<li>\u8fd8\u6709\u66f4\u591a&#8230;<\/li>\n<\/ul>\n\n\n\n<p>OpenDiLoCo \u662f\u6211\u4eec\u7684\u7814\u7a76\u5de5\u4f5c\u4e4b\u4e00\uff0c\u65e8\u5728\u901a\u8fc7\u4fc3\u8fdb\u5168\u7403\u591a\u4e2a\u8fde\u63a5\u4e0d\u826f\u8bbe\u5907\u7684\u9ad8\u6548\u8bad\u7ec3\u6765\u514b\u670d\u8fd9\u4e9b\u6311\u6218\u4e2d\u7684\u7b2c\u4e00\u4e2a\u6311\u6218\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e3b\u8981\u8d21\u732e<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u590d\u5236\u548c\u6269\u5c55<\/strong>\uff1a\u6211\u4eec\u5df2\u7ecf\u6210\u529f\u5730\u91cd\u73b0\u4e86\u539f\u59cb\u7684 DiLoCo \u5b9e\u9a8c\uff0c\u5e76\u5c06\u5176\u6269\u5c55\u5230\u5341\u4ebf\u53c2\u6570\u6a21\u578b\u89c4\u6a21\u3002<\/li>\n\n\n\n<li><strong>\u5f00\u6e90\u5b9e\u73b0<\/strong>\uff1a\u6211\u4eec\u6b63\u5728\u53d1\u5e03\u4e00\u4e2a\u57fa\u4e8e\u00a0<a href=\"https:\/\/github.com\/learning-at-home\/hivemind\">Hivemind<\/a>\u00a0\u5e93\u7684\u53ef\u6269\u5c55\u5b9e\u73b0\uff0c\u4f7f\u5e7f\u6cdb\u7684\u5f00\u53d1\u4eba\u5458\u548c\u7814\u7a76\u4eba\u5458\u90fd\u53ef\u4ee5\u8fdb\u884c\u53bb\u4e2d\u5fc3\u5316\u57f9\u8bad\u3002\u6211\u4eec\u7684\u6846\u67b6\u901a\u8fc7\u4e0e PyTorch FSDP \u7684\u96c6\u6210\uff0c\u4f7f\u5355\u4e2a DiLoCo \u5de5\u4f5c\u4eba\u5458\u80fd\u591f\u6269\u5c55\u5230\u6570\u767e\u53f0\u673a\u5668\u3002<\/li>\n\n\n\n<li><strong>\u5168\u7403\u53bb\u4e2d\u5fc3\u5316\u8bad\u7ec3<\/strong>\uff1a\u6211\u4eec\u901a\u8fc7\u5728\u4e24\u5927\u6d32\u548c\u4e09\u4e2a\u56fd\u5bb6\/\u5730\u533a\u8bad\u7ec3\u6a21\u578b\uff0c\u5b9e\u73b0\u4e86 90-95% \u7684\u8ba1\u7b97\u5229\u7528\u7387\uff0c\u5c55\u793a\u4e86 OpenDiLoCo \u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684\u6f5c\u529b\u3002<\/li>\n\n\n\n<li><strong>\u6548\u7387\u6d1e\u5bdf<\/strong>\uff1a\u6211\u4eec\u7684\u6d88\u878d\u7814\u7a76\u4e3a\u7b97\u6cd5\u7684\u53ef\u6269\u5c55\u6027\u548c\u8ba1\u7b97\u6548\u7387\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u89c1\u89e3\uff0c\u4e3a\u672a\u6765\u7684\u6539\u8fdb\u94fa\u5e73\u4e86\u9053\u8def\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">DiLoCo<\/h2>\n\n\n\n<p>Google DeepMind\u6700\u8fd1\u7684\u5de5\u4f5c\u5f15\u5165\u4e86\u4e00\u79cd\u65b9\u6cd5\uff0c\u53ef\u4ee5\u5728\u8fde\u63a5\u4e0d\u826f\u7684\u8bbe\u5907\u5b64\u5c9b\u4e0a\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u5728\u8fd9\u4e9b\u4e0d\u540c\u7684\u5c9b\u5c7f\u4e0a\u8fdb\u884c\u6570\u636e\u5e76\u884c\u8bad\u7ec3\uff0c\u53ea\u9700\u8981\u6bcf 500 \u6b65\u540c\u6b65\u4e00\u6b21\u4f2a\u68af\u5ea6\u3002<\/p>\n\n\n\n<p>DiLoCo \u5f15\u5165\u4e86\u4e00\u79cd\u5185\u5916\u4f18\u5316\u7b97\u6cd5\uff0c\u5141\u8bb8\u672c\u5730\u548c\u5168\u5c40\u66f4\u65b0\u3002\u6bcf\u4e2a\u5de5\u4f5c\u7ebf\u7a0b\u4f7f\u7528\u672c\u5730&nbsp;<a href=\"https:\/\/arxiv.org\/abs\/1711.05101\">AdamW<\/a>&nbsp;\u4f18\u5316\u5668\uff08\u5185\u90e8\u4f18\u5316\uff09\u72ec\u7acb\u5730\u591a\u6b21\u66f4\u65b0\u5176\u6743\u91cd\u3002\u6bcf ~500 \u6b21\u66f4\u65b0\uff0c\u8be5\u7b97\u6cd5\u5c31\u4f1a\u4f7f\u7528&nbsp;<a href=\"https:\/\/www.semanticscholar.org\/paper\/A-method-for-solving-the-convex-programming-problem-Nesterov\/8d3a318b62d2e970122da35b2a2e70a5d12cc16f\">Nesterov \u52a8\u91cf<\/a>\u4f18\u5316\u5668\u6267\u884c<em>\u5916\u90e8\u4f18\u5316<\/em>\uff0c\u8be5\u4f18\u5316\u5668\u540c\u6b65\u6240\u6709\u5de5\u4f5c\u4eba\u5458\u7684\u4f2a\u68af\u5ea6\uff08\u6240\u6709\u5c40\u90e8\u68af\u5ea6\u7684\u603b\u548c\uff09\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u65b9\u6cd5\u5927\u5927\u964d\u4f4e\u4e86\u901a\u4fe1\u9891\u7387\uff08\u6700\u591a 500 \u6b21\uff09\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u5e26\u5bbd\u8981\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"2764\" height=\"932\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1.png\" alt=\"\" class=\"wp-image-4292\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1.png 2764w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-300x101.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-1024x345.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-768x259.png 768w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-1536x518.png 1536w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-2048x691.png 2048w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-1320x445.png 1320w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1-600x202.png 600w\" sizes=\"auto, (max-width: 2764px) 100vw, 2764px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">OpenDiLoCo<\/h2>\n\n\n\n<p>\u4e3a\u4e86\u4fc3\u8fdb\u5728\u8fd9\u4e2a\u6709\u524d\u9014\u7684\u7814\u7a76\u65b9\u5411\u4e0a\u7684\u5408\u4f5c\uff0c\u4f7f\u4eba\u5de5\u667a\u80fd\u6c11\u4e3b\u5316\uff0c\u6211\u4eec\u5728\u5f00\u6e90\u8bb8\u53ef\u4e0b\u53d1\u5e03\u4e86 OpenDiLoCo \u4ee3\u7801\uff1a<a href=\"https:\/\/github.com\/PrimeIntellect-ai\/OpenDiLoCo\">https:\/\/github.com\/PrimeIntellect-ai\/OpenDiLoCo<\/a>\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u7684\u5b9e\u73b0\u5efa\u7acb\u5728&nbsp;<a href=\"https:\/\/github.com\/learning-at-home\/hivemind\">Hivemind \u5e93<\/a>\u4e4b\u4e0a\u3002Hivemind \u6ca1\u6709\u4f7f\u7528&nbsp;<em>torch.distributed<\/em>&nbsp;\u8fdb\u884c\u5de5\u4f5c\u7ebf\u7a0b\u901a\u4fe1\uff0c\u800c\u662f\u5229\u7528\u5206\u5e03\u5728\u6bcf\u4e2a\u5de5\u4f5c\u7ebf\u7a0b\u4e0a\u7684<a href=\"https:\/\/en.wikipedia.org\/wiki\/Distributed_hash_table\">\u5206\u5e03\u5f0f\u54c8\u5e0c\u8868 \uff08DHT\uff09<\/a>&nbsp;\u6765\u901a\u4fe1\u5143\u6570\u636e\u5e76\u540c\u6b65\u5b83\u4eec\u3002\u6b64 DHT \u662f\u4f7f\u7528\u5f00\u6e90&nbsp;<a href=\"https:\/\/github.com\/libp2p\/libp2p\">libp2p \u9879\u76ee<\/a>\u5b9e\u73b0\u7684\u3002\u6211\u4eec\u5229\u7528 Hivemind \u8fdb\u884c DiLoCo worker \u4e4b\u95f4\u7684\u8282\u70b9\u95f4\u901a\u4fe1\uff0c\u5e76\u5229\u7528 PyTorch FSDP \u8fdb\u884c DiLoCo Worker \u5185\u7684\u8282\u70b9\u5185\u901a\u4fe1\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u4e0e Hivemind \u7684\u96c6\u6210\u4e3a DiLoCo \u63d0\u4f9b\u4e86\u771f\u5b9e\u4e16\u754c\u7684\u53bb\u4e2d\u5fc3\u5316\u8bad\u7ec3\u8bbe\u7f6e\uff0c\u4f7f\u5176\u8bb8\u591a\u56fa\u6709\u5c5e\u6027\u53ef\u7528\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8d44\u6e90\u7684\u5f00\/\u5173\u659c\u5761\uff1a<\/strong>\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u53ef\u7528\u8ba1\u7b97\u91cf\u53ef\u80fd\u4f1a\u6709\u6240\u4e0d\u540c\uff0c\u65b0\u8bbe\u5907\u548c\u96c6\u7fa4\u4f1a\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u52a0\u5165\u548c\u79bb\u5f00\u3002<\/li>\n\n\n\n<li><strong>\u5bb9\u9519\uff1a<\/strong>\u5bf9\u4e8e\u5206\u6563\u5f0f\u57f9\u8bad\uff0c\u67d0\u4e9b\u8bbe\u5907\u53ef\u80fd\u4e0d\u5982\u5176\u4ed6\u8bbe\u5907\u53ef\u9760\u3002\u901a\u8fc7 Hivemind \u7684\u5bb9\u9519\u8bad\u7ec3\uff0c\u8bbe\u5907\u53ef\u4ee5\u5728\u4e0d\u505c\u6b62\u8bad\u7ec3\u8fc7\u7a0b\u7684\u60c5\u51b5\u4e0b\u968f\u65f6\u53d8\u5f97\u4e0d\u53ef\u7528\u3002<\/li>\n\n\n\n<li><strong>\u70b9\u5bf9\u70b9\uff1a<\/strong>\u6ca1\u6709\u4e3b\u8282\u70b9\u3002\u6240\u6709\u901a\u4fe1\u90fd\u662f\u4ee5\u70b9\u5bf9\u70b9\u7684\u65b9\u5f0f\u5b8c\u6210\u7684\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e3b\u8981\u7ed3\u679c<\/h2>\n\n\n\n<p>\u4f5c\u4e3a\u7b2c\u4e00\u6b65\uff0c\u6211\u4eec\u590d\u5236\u4e86DiLoCO\u7684\u4e3b\u8981\u5b9e\u9a8c\u7ed3\u679c\u3002\u6211\u4eec\u4f7f\u7528 C4 \u6570\u636e\u96c6\u5728\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e0a\u8bad\u7ec3\u4e86\u4e00\u4e2a\u5177\u6709 1.5 \u4ebf\u4e2a\u53c2\u6570\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u53d1\u73b0\uff0c\u5c3d\u7ba1\u901a\u4fe1\u8981\u6c42\u964d\u4f4e\u4e86 500 \u500d\uff0c\u4f46\u5177\u6709 8 \u4e2a\u526f\u672c\u7684 DiLoCo \u660e\u663e\u4f18\u4e8e\u6ca1\u6709\u4efb\u4f55\u526f\u672c\u7684\u57fa\u7ebf\uff0c\u5e76\u4e14\u5728\u76f8\u540c\u7684\u8ba1\u7b97\u9884\u7b97\u4e0b\u4e0e\u66f4\u5f3a\u7684\u57fa\u7ebf\u7684\u6027\u80fd\u76f8\u5339\u914d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"4097\" height=\"3277\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2.png\" alt=\"\" class=\"wp-image-4293\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2.png 4097w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-300x240.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-1024x819.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-768x614.png 768w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-1536x1229.png 1536w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-2048x1638.png 2048w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-1320x1056.png 1320w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-2-600x480.png 600w\" sizes=\"auto, (max-width: 4097px) 100vw, 4097px\" \/><figcaption class=\"wp-element-caption\">\u4f7f\u7528 8 \u4e2a DiLoCo \u5de5\u4f5c\u4eba\u5458\u8fdb\u884c 1.5 \u4ebf\u53c2\u6570 Llama \u6a21\u578b\u9884\u8bad\u7ec3\u4ea7\u751f\u7684\u56f0\u60d1\u5ea6\u660e\u663e\u4f4e\u4e8e\u6ca1\u6709 DiLoCo \u7684\u57fa\u7ebf\uff0c\u751a\u81f3\u4e0e\u57fa\u7ebf\u76f8\u6bd4\uff0c\u4f7f\u7528\u76f8\u540c\u8ba1\u7b97\u9884\u7b97\u7684\u6279\u91cf\u5927\u5c0f\u5927 8 \u500d\uff0c\u540c\u65f6\u901a\u4fe1\u91cf\u51cf\u5c11 500 \u500d\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u200d<\/p>\n\n\n\n<p>\u9664\u4e86\u590d\u5236\u7ed3\u679c\u5916\uff0c\u6211\u4eec\u8fd8\u5bf9 DiLoCo \u8fdb\u884c\u4e86\u51e0\u9879\u6d88\u878d\u7814\u7a76\uff0c\u91cd\u70b9\u5173\u6ce8\u7b97\u6cd5\u7684\u53ef\u6269\u5c55\u6027\u4e0e\u5de5\u4f5c\u7ebf\u7a0b\u6570\u91cf\u548c\u8ba1\u7b97\u6548\u7387\u3002\u6211\u4eec\u8fd8\u8bc1\u660e\uff0c\u5728FP16\u4e2d\uff0cDiLoCo\u4f2a\u68af\u5ea6\u53ef\u4ee5\u6709\u6548\u5730\u5168\u90e8\u8fd8\u539f\uff0c\u800c\u4e0d\u4f1a\u964d\u4f4e\u4efb\u4f55\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u6709\u5173\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u67e5\u770b<a href=\"https:\/\/arxiv.org\/abs\/2407.07852\">arXiv<\/a>\u4e0a\u7684\u8bba\u6587\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5c06 DiLoCo \u6269\u5c55\u5230\u5341\u4ebf\u4e2a\u53c2\u6570\u6a21\u578b<\/h2>\n\n\n\n<p>DeepMind \u6700\u521d\u7684 DiLoCo \u8bba\u6587\u53ea\u8bd5\u9a8c\u4e86\u591a\u8fbe 4 \u4ebf\u4e2a\u53c2\u6570\u7684\u6a21\u578b\u5927\u5c0f\u3002\u5728\u6211\u4eec\u7684\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5c06\u8be5\u65b9\u6cd5\u6269\u5c55\u5230\u5177\u6709 11 \u4ebf\u4e2a\u53c2\u6570\u7684\u6a21\u578b\u3002\u6211\u4eec\u91c7\u7528\u4e0e&nbsp;<a href=\"https:\/\/github.com\/jzhang38\/TinyLlama\">TinyLlama<\/a>&nbsp;\u76f8\u540c\u7684\u8d85\u53c2\u6570\uff0c\u5e76\u4f7f\u7528\u5305\u542b 800 \u4e07\u4e2a\u4ee4\u724c\u7684\u603b\u6279\u5927\u5c0f\uff08\u6279\u5927\u5c0f\u4e3a 8192\uff0c\u5e8f\u5217\u957f\u5ea6\u4e3a 1024\uff09\u3002\u7531\u4e8e\u4e0e\u6211\u4eec\u4e4b\u524d\u7684\u5b9e\u9a8c\u76f8\u6bd4\uff0c\u6279\u91cf\u5927\u5c0f\u589e\u52a0\u4e86 4 \u500d\uff0c\u56e0\u6b64\u6211\u4eec\u51b3\u5b9a\u53ea\u4e3a\u8fd9\u4e2a\u5b9e\u9a8c\u8bad\u7ec3 44k \u4e2a\u6b65\u9aa4\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u7ed3\u679c\u4e0e\u4e24\u4e2a\u57fa\u7ebf\u8fdb\u884c\u4e86\u6bd4\u8f83\uff1a\u4e00\u4e2a\u6ca1\u6709 DiLoCo \u548c\u6ca1\u6709\u526f\u672c\u7684\u5f31\u57fa\u7ebf\uff0c\u4ee5\u53ca\u4f7f\u7528\u5177\u6709\u6570\u636e\u5e76\u884c\u6027\u7684 4\u00d7 \u5927\u6279\u91cf\u7684\u66f4\u5f3a\u57fa\u7ebf\u3002<\/p>\n\n\n\n<p>\u5f53\u4f7f\u7528 500 \u7684\u5c40\u90e8\u6b65\u957f\uff08\u6bcf 500 \u6b65\u5728\u5de5\u4eba\u4e4b\u95f4\u540c\u6b65\uff09\u65f6\uff0c\u7c7b\u4f3c\u4e8e 1.5 \u4ebf\u53c2\u6570\u6a21\u578b\u7684\u5b9e\u9a8c\uff0c\u6211\u4eec\u5728\u8bad\u7ec3\u7684\u65e9\u671f\u9636\u6bb5\u89c2\u5bdf\u5230\u6b21\u4f18\u6536\u655b\u3002\u8bad\u7ec3\u52a8\u6001\u5728\u8bad\u7ec3\u7684\u540e\u671f\u9636\u6bb5\u6709\u6240\u6539\u5584\uff0c\u5982\u679c\u6211\u4eec\u8fdb\u4e00\u6b65\u8bad\u7ec3\u5230 88k \u6b65\uff0c\u53ef\u80fd\u4f1a\u7b26\u5408\u6211\u4eec\u7684\u57fa\u7ebf\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u4e00\u9879\u5c40\u90e8\u6b65\u957f\u4e3a 125 \u7684\u5b9e\u9a8c\u3002\u5728\u8fd9\u79cd\u5236\u5ea6\u4e0b\uff0c\u5728\u8bad\u7ec3\u7684\u65e9\u671f\u9636\u6bb5\uff0c\u8bad\u7ec3\u52a8\u6001\u5b9e\u9645\u4e0a\u66f4\u597d\u3002\u5177\u6709 125 \u4e2a\u672c\u5730\u6b65\u9aa4\u7684 DiLoCo \u8fd0\u884c\u51e0\u4e4e\u4e0e\u76f8\u540c\u8ba1\u7b97\u9884\u7b97\u4e0b\u7684\u66f4\u5f3a\u57fa\u7ebf\u7684\u6027\u80fd\u76f8\u5339\u914d\uff0c\u800c\u901a\u4fe1\u91cf\u51cf\u5c11\u4e86 125 \u500d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"4097\" height=\"3277\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3.png\" alt=\"\" class=\"wp-image-4294\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3.png 4097w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-300x240.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-1024x819.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-768x614.png 768w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-1536x1229.png 1536w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-2048x1638.png 2048w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-1320x1056.png 1320w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-3-600x480.png 600w\" sizes=\"auto, (max-width: 4097px) 100vw, 4097px\" \/><figcaption class=\"wp-element-caption\"><strong>1.1B \u7f29\u653e\u5b9e\u9a8c\uff1a<\/strong>\u6bd4\u8f83 OpenDiLoCo \u7684 1.1B \u8bad\u7ec3\uff0c\u5176\u4e2d 4 \u540d\u5de5\u4f5c\u4eba\u5458\u6bcf 500 \u4e2a\u672c\u5730\u6b65\u9aa4\u548c\u6bcf 125 \u4e2a\u672c\u5730\u6b65\u9aa4\u4e0e\u4e24\u4e2a\u57fa\u7ebf\u540c\u6b65\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u867d\u7136\u6211\u4eec\u8bc1\u660e\u4e86 DiLoCo \u5728\u5341\u4ebf\u53c2\u6570\u89c4\u6a21\u4e0b\u5de5\u4f5c\uff0c\u4f46\u6211\u4eec\u8ba4\u4e3a\u8fd8\u9700\u8981\u8fdb\u4e00\u6b65\u7684\u5de5\u4f5c\u624d\u80fd\u4f7f\u5176\u5728\u66f4\u5927\u7684\u6279\u91cf\u5927\u5c0f\u548c\u589e\u52a0\u7684\u672c\u5730\u6b65\u9aa4\u4e2d\u6709\u6548\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5168\u7403\u5206\u5e03\u5f0f\u8bad\u7ec3\u8bbe\u7f6e<\/h2>\n\n\n\n<p>\u4e3a\u4e86\u5c55\u793a OpenDiLoCo \u5728\u4e0d\u540c\u5927\u6d32\u8fdb\u884c\u5206\u6563\u8bad\u7ec3\u7684\u529f\u80fd\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u56db\u4e2a DiLoCo \u5de5\u4f5c\u4eba\u5458\uff0c\u6bcf\u4e2a\u5de5\u4f5c\u4eba\u5458\u6709\u516b\u4e2a H100 GPU\uff0c\u5206\u522b\u4f4d\u4e8e\u52a0\u62ff\u5927\u3001\u82ac\u5170\u548c\u7f8e\u56fd\u7684\u4e24\u4e2a\u4e0d\u540c\u5dde\u3002\u56fe\u4e2d\u663e\u793a\u4e86\u5de5\u4f5c\u7ebf\u7a0b\u4e4b\u95f4\u7684\u7f51\u7edc\u5e26\u5bbd\uff0c\u5728127\u5230935 Mbit\/s\u4e4b\u95f4\u53d8\u5316\u3002\u6211\u4eec\u7528 500 \u4e2a\u5c40\u90e8\u6b65\u9aa4\u8bad\u7ec3\u4e86\u6211\u4eec\u7684 1.1B \u53c2\u6570\u6a21\u578b\uff0c\u5728 FP16 \u4e2d\u68af\u5ea6\u5168\u90e8\u51cf\u5c11\u3002\u7531\u4e8e\u672c\u5730\u6b65\u9aa4\u6570\u91cf\u4f17\u591a\uff0c\u56db\u540d\u5de5\u4f5c\u4eba\u5458\u72ec\u7acb\u8fd0\u884c\u4e86\u5927\u7ea6 67.5 \u5206\u949f\uff0c\u7136\u540e\u8fdb\u884c\u68af\u5ea6\u5e73\u5747\u901a\u4fe1\u3002\u5bf9\u4e8e\u5916\u90e8\u4f18\u5316\u5668\u6b65\u9aa4\uff0c\u6211\u4eec\u7684\u5b9e\u9a8c\u663e\u793a\uff0c\u5de5\u4f5c\u7ebf\u7a0b\u4e4b\u95f4\u7684\u5e73\u5747\u5168\u7f29\u51cf\u65f6\u95f4\u4e3a 300 \u79d2\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1805\" height=\"855\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image.png\" alt=\"\" class=\"wp-image-4291\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image.png 1805w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-300x142.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1024x485.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-768x364.png 768w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1536x728.png 1536w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-1320x625.png 1320w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-600x284.png 600w\" sizes=\"auto, (max-width: 1805px) 100vw, 1805px\" \/><figcaption class=\"wp-element-caption\">\u5458\u5de5\u4e4b\u95f4\u7684\u7f51\u7edc\u5e26\u5bbd\uff1a\u4f7f\u7528iperf\u6d4b\u91cf\u7684\u5206\u6563\u5f0f\u57f9\u8bad\u8bbe\u7f6e\u4e2d\u56db\u4e2a\u5458\u5de5\u4e4b\u95f4\u7684\u5e73\u5747\u53cc\u5411\u7f51\u7edc\u5e26\u5bbd\uff08\u4ee5Mbit\/s\u4e3a\u5355\u4f4d\uff09\u3002\u8fd9\u4e9b\u8282\u70b9\u4f4d\u4e8e\u4e09\u4e2a\u4e0d\u540c\u7684\u56fd\u5bb6\uff0c\u7531\u4e0d\u540c\u7684\u4e91\u63d0\u4f9b\u5546\u6258\u7ba1\uff1a\u52a0\u62ff\u5927\uff08Hyperstack\uff09\u3001\u82ac\u5170\uff08DataCrunch\uff09\u3001\u7f8e\u56fd\u5fb7\u514b\u8428\u65af\u5dde\uff08Voltage Park\uff09\u3001\u7f8e\u56fd\u7279\u62c9\u534e\u5dde\uff08Runpod\uff09\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u7531\u4e8e DiLoCo \u663e\u8457\u7f29\u77ed\u4e86\u901a\u4fe1\u65f6\u95f4\uff0c\u56e0\u6b64 all-reduce \u74f6\u9888\u4ec5\u5360\u8bad\u7ec3\u65f6\u95f4\u7684 6.9%\uff0c\u5bf9\u6574\u4f53\u8bad\u7ec3\u901f\u5ea6\u7684\u5f71\u54cd\u6700\u5c0f\u3002<\/p>\n\n\n\n<p>\u5728\u6211\u4eec\u7684\u573a\u666f\u4e2d\uff0c\u6700\u5feb\u7684\u5de5\u4f5c\u4eba\u5458\u5c06\u82b1\u8d39\u989d\u5916\u7684\u8bad\u7ec3\u65f6\u95f4\u3002\u5728\u672a\u6765\u7684\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5c06\u901a\u8fc7<a href=\"https:\/\/arxiv.org\/abs\/2401.09135v1\">\u5728\u5f02\u6b65\u73af\u5883\u4e2d\u63a2\u7d22 DiLoCo<\/a>&nbsp;\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u8fd0\u884c OpenDiLoCo<\/h2>\n\n\n\n<p>\u8fd0\u884c\u4ee3\u7801\u5f88\u7b80\u5355\u3002\u552f\u4e00\u7684\u8981\u6c42\u662f\u53ef\u4ee5\u8bbf\u95ee\u81f3\u5c11\u4e24\u4e2a GPU\uff0c\u5b83\u4eec\u4e0d\u5fc5\u4f4d\u4e8e\u540c\u4e00\u4f4d\u7f6e\u3002\u8bbe\u7f6e\u73af\u5883\u540e\uff0c\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u521b\u5efa\u521d\u59cb DHT \u8282\u70b9\uff1a<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \" >python ..\/hivemind_source\/hivemind\/hivemind_cli\/run_dht.py\n\u00a0 \u00a0--identity_path fixed_private_key.pem\n\u00a0 \u00a0--host_maddrs \/ip4\/0.0.0.0\/tcp\/30001<\/pre><\/div>\n\n\n\n<p>\u5728\u53e6\u4e00\u4e2a\u7ec8\u7aef\u4e2d\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u542f\u52a8 DiLoCo worker\uff0c\u786e\u4fdd\u5c06 \u548c \u9002\u5f53\u5730\u8bbe\u7f6e\uff1a<code>PEERNUM_DILOCO_WORKERSWORLD_RANK<\/code><\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \" >export PEER=\/ip4\/192.168.100.20\/tcp\/30001\/p2p\/Qmbh7opLJxFCtY22XqwETuo6bnWqijs76YXz7D69MBWEuZ\n# change the IP above to your public IP if using across nodes connected via internet\nexport NUM_DILOCO_WORKERS=4\nexport WORLD_RANK=0\n\ntorchrun --nproc_per_node=8 \\\\\n\u00a0 \u00a0train_fsdp.py \\\\\n\u00a0 \u00a0--per-device-train-batch-size 16 \\\\\n\u00a0 \u00a0--total-batch-size 2048 \\\\\n\u00a0 \u00a0--total-steps 88_000 \\\\\n\u00a0 \u00a0--project OpenDiLoCo \\\\\n\u00a0 \u00a0--lr 4e-4 \\\\\n\u00a0 \u00a0--model-name-or-path PrimeIntellect\/llama-1b-fresh \\\\\n\u00a0 \u00a0--warmup-steps 1000 \\\\\n\u00a0 \u00a0--hv.averaging_timeout 1800 \\\\\n\u00a0 \u00a0--hv.skip_load_from_peers \\\\\n\u00a0 \u00a0--hv.local_steps 500 \\\\\n\u00a0 \u00a0--hv.initial-peers $PEER \\\\\n\u00a0 \u00a0--hv.galaxy-size $NUM_DILOCO_WORKERS \\\\\n\u00a0 \u00a0--hv.world-rank $WORLD_RANK \\\\\n\u00a0 \u00a0--checkpoint_interval 500 \\\\\n\u00a0 \u00a0--checkpoint-path 1b_diloco_ckpt<\/pre><\/div>\n\n\n\n<p>\u60a8\u53ef\u4ee5\u5728 GitHub \u5b58\u50a8\u5e93\u7684&nbsp;<a href=\"https:\/\/github.com\/PrimeIntellect-ai\/OpenDiLoCo\/blob\/main\/README.md\">README<\/a>&nbsp;\u4e2d\u627e\u5230\u6709\u5173\u8fd0\u884c OpenDiLoCo \u7684\u66f4\u591a\u4fe1\u606f\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5728 PI \u8ba1\u7b97\u5e73\u53f0\u4e0a\u8fd0\u884c OpenDiLoCo<\/h2>\n\n\n\n<p>\u8bbe\u7f6e\u5168\u5c40\u7f16\u6392\u5c42\u4ee5\u8fd0\u884c DiLoCo \u8bad\u7ec3\u8fd0\u884c\u4ecd\u7136\u975e\u5e38\u5177\u6709\u6311\u6218\u6027\u3002\u5f97\u76ca\u4e8e\u6211\u4eec\u9884\u6784\u5efa\u7684 OpenDiLoCo docker \u6620\u50cf\uff0c\u6211\u4eec\u7684&nbsp;<a href=\"https:\/\/app.primeintellect.ai\/\">PI \u8ba1\u7b97\u5e73\u53f0<\/a>\u4f7f\u8fd9\u53d8\u5f97\u66f4\u52a0\u5bb9\u6613\u3002\u8be5\u6620\u50cf\u9884\u88c5\u4e86\u6240\u6709\u4f9d\u8d56\u9879\uff0c\u5141\u8bb8\u4eba\u4eec\u8f7b\u677e\u751f\u6210 DiLoCo \u5de5\u4f5c\u7ebf\u7a0b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"2934\" height=\"904\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4.png\" alt=\"\" class=\"wp-image-4295\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4.png 2934w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-300x92.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-1024x316.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-768x237.png 768w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-1536x473.png 1536w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-2048x631.png 2048w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-1320x407.png 1320w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/07\/image-4-600x185.png 600w\" sizes=\"auto, (max-width: 2934px) 100vw, 2934px\" \/><\/figure>\n\n\n\n<p>\u5728\u672a\u6765\u7684\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5728\u8ba1\u7b97\u5e73\u53f0\u4e2d\u6784\u5efa\u4e00\u4e2a\u96c6\u6210\u7684\u5f00\u6e90\u5806\u6808\uff0c\u4e3a\u8de8\u591a\u4e2a\u96c6\u7fa4\u7684\u7f16\u6392\u3001\u6548\u7387\u4f18\u5316\u3001\u5904\u7406\u8282\u70b9\u6545\u969c\u3001\u57fa\u7840\u8bbe\u65bd\u76d1\u63a7\u7b49\u63d0\u4f9b\u6d41\u7545\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u8bba\u548c\u672a\u6765\u65b9\u5411<\/h2>\n\n\n\n<p>\u6211\u4eec\u6210\u529f\u518d\u73b0\u4e86DiLoCo\u7684\u4e3b\u8981\u5b9e\u9a8c\u7ed3\u679c\uff0c\u5c06\u8be5\u65b9\u6cd5\u7684\u53c2\u6570\u5927\u5c0f\u7f29\u653e\u5230\u539f\u59cb\u5de5\u4f5c\u7684\u4e09\u500d\uff0c\u5e76\u6f14\u793a\u4e86\u5176\u5728\u771f\u5b9e\u4e16\u754c\u7684\u5206\u6563\u8bad\u7ec3\u73af\u5883\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u672a\u6765\u7684\u5de5\u4f5c\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u5c06 DiLoCo \u6269\u5c55\u5230\u66f4\u591a\u5206\u5e03\u5f0f\u5de5\u4f5c\u7ebf\u7a0b\u7684\u66f4\u5927\u6a21\u578b\u3002\u4e00\u4e9b\u6709\u8da3\u7684\u65b9\u5411\u5305\u62ec\u53ef\u4ee5\u63d0\u9ad8\u7a33\u5b9a\u6027\u548c\u6536\u655b\u901f\u5ea6\u7684\u6a21\u578b\u5408\u5e76\u6280\u672f\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u5b9e\u73b0\u5f02\u6b65\u6267\u884c\u6743\u91cd\u5e73\u5747\u901a\u4fe1\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u51cf\u5c11\u8ba1\u7b97\u7a7a\u95f2\u65f6\u95f4\uff0c\u5c06\u5b83\u4eec\u4e0e\u4e0b\u4e00\u4e2a\u5916\u90e8\u4f18\u5316\u6b65\u9aa4\u7684\u8ba1\u7b97\u4ea4\u7ec7\u5728\u4e00\u8d77\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5bf9\u8fd9\u9879\u6280\u672f\u7684\u76f4\u63a5\u5b9e\u9645\u5e94\u7528\u611f\u5230\u5174\u594b\uff0c\u5e76\u671f\u5f85\u5728\u4e0d\u4e45\u7684\u5c06\u6765\u4e3a\u6211\u4eec\u603b\u4f53\u89c4\u5212\u7684\u7b2c\u4e09\u90e8\u5206\u5efa\u7acb\u5b83\uff1a\u5728\u8bed\u8a00\u3001\u4ee3\u7406\u3001\u4ee3\u7801\u548c\u79d1\u5b66\u7b49\u9ad8\u5f71\u54cd\u529b\u9886\u57df\u534f\u4f5c\u8bad\u7ec3\u548c\u8d21\u732e\u5f00\u653e AI \u6a21\u578b\uff0c\u4ee5\u5b9e\u73b0 AI \u6a21\u578b\u7684\u96c6\u4f53\u6240\u6709\u6743\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u52a0\u5165\u6211\u4eec\uff0c\u5171\u5efa\u4eba\u5de5\u667a\u80fd\u7684\u5f00\u653e\u672a\u6765<\/h2>\n\n\n\n<p>\u5851\u9020\u4eba\u5de5\u667a\u80fd\u672a\u6765\u7684\u6743\u529b\u4e0d\u5e94\u8be5\u96c6\u4e2d\u5728\u5c11\u6570\u4eba\u624b\u4e2d\uff0c\u800c\u5e94\u8be5\u5411\u4efb\u4f55\u6709\u80fd\u529b\u505a\u51fa\u8d21\u732e\u7684\u4eba\u5f00\u653e\u3002\u6211\u4eec\u8bda\u9080\u60a8\u52a0\u5165\u6211\u4eec\uff0c\u5171\u540c\u6784\u5efa\u4e00\u4e2a\u66f4\u52a0\u5206\u5e03\u5f0f\u548c\u6709\u5f71\u54cd\u529b\u7684\u4eba\u5de5\u667a\u80fd\u672a\u6765\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c\u60a8\u96c4\u5fc3\u52c3\u52c3\u5e76\u5e0c\u671b\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u8bf7\u7533\u8bf7\u6211\u4eec\u7684<a href=\"https:\/\/jobs.ashbyhq.com\/PrimeIntellect\">\u7a7a\u7f3a\u804c\u4f4d<\/a>\u3002<\/li>\n\n\n\n<li><a href=\"https:\/\/www.primeintellect.ai\/blog\/fast-compute-grants\">\u5c31\u6211\u4eec\u7684 AI \u6a21\u578b\u8ba1\u5212<\/a>\u548c<a href=\"https:\/\/github.com\/PrimeIntellect-ai\/OpenDiloco\">\u5f00\u6e90\u6846\u67b6<\/a>\u8fdb\u884c\u534f\u4f5c\u3002<\/li>\n\n\n\n<li>\u5728\u6700\u5148\u8fdb\u7684 AI \u6a21\u578b\u4e2d\u8d21\u732e\u3001\u8ba1\u7b97\u548c\u8d62\u5f97\u6240\u6709\u6743\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6211\u4eec\u7684 OpenDiLoCo \u5de5\u4f5c\u4e5f\u88ab ICML \u7684&nbsp;<a href=\"https:\/\/es-fomo.com\/\">ES-FoMo<\/a>&nbsp;\u7814\u8ba8\u4f1a\u6240\u63a5\u53d7\u3002\u5982\u679c\u60a8\u8981\u6765\u7ef4\u4e5f\u7eb3\uff0c\u8bf7\u8054\u7cfb\u6211\u4eec\uff01<\/p>\n\n\n\n<p>\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/www.primeintellect.ai\/blog\/opendiloco\">https:\/\/www.primeintellect.ai\/blog\/opendiloco<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9686\u91cd\u63a8\u51fa OpenDiLoCo\uff0c\u8fd9\u662f DeepMind \u5206\u5e03\u5f0f\u4f4e\u901a\u4fe1 \uff08DiLoCo\uff09 \u65b9\u6cd5\u7684\u5f00\u6e90\u5b9e\u73b0\u548c\u6269\u5c55\uff0c [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[444,445,443,442],"tags":[242],"class_list":["post-4289","post","type-post","status-publish","format-standard","hentry","category-ai","category-ainews","category-llm","category-llms","tag-chatgpt"],"views":2187,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4289","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4289"}],"version-history":[{"count":3,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4289\/revisions"}],"predecessor-version":[{"id":4298,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4289\/revisions\/4298"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4289"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4289"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4289"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}