{"id":1071024,"date":"2023-02-01T01:04:00","date_gmt":"2023-02-01T09:04:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&#038;p=1071024"},"modified":"2024-09-25T03:29:18","modified_gmt":"2024-09-25T10:29:18","slug":"torchscale","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/torchscale\/","title":{"rendered":"\u8ba9\u5929\u4e0b\u6ca1\u6709\u96be\u8bad\u7ec3\u7684\u5927\u6a21\u578b\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5f00\u6e90TorchScale"},"content":{"rendered":"\n<p>\u7f16\u8005\u6309\uff1a\u8fd1\u671f\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ece\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u7406\u8bba\u51fa\u53d1\uff0c\u7814\u53d1\u5e76\u63a8\u51fa\u4e86 TorchScale \u5f00\u6e90\u5de5\u5177\u5305\u3002TorchScale \u5de5\u5177\u5305\u901a\u8fc7\u91c7\u7528 DeepNet\u3001Magneto \u548c X-MoE \u7b49\u6700\u5148\u8fdb\u7684\u5efa\u6a21\u6280\u672f\uff0c\u53ef\u4ee5\u5e2e\u52a9\u7814\u7a76\u548c\u5f00\u53d1\u4eba\u5458\u63d0\u9ad8\u5efa\u6a21\u7684\u901a\u7528\u6027\u548c\u6574\u4f53\u6027\u80fd\uff0c\u786e\u4fdd\u8bad\u7ec3\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u53ca\u6548\u7387\uff0c\u5e76\u5141\u8bb8\u4ee5\u4e0d\u540c\u7684\u6a21\u578b\u5927\u5c0f\u6269\u5c55 Transformer \u7f51\u7edc\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u5982\u4eca\uff0c\u5728\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\uff08CV\uff09\u3001\u8bed\u97f3\u3001\u591a\u6a21\u6001\u6a21\u578b\u548c AI for Science \u7b49\u9886\u57df\u7814\u7a76\u4e2d\uff0cTransformer \u5df2\u7ecf\u6210\u4e3a\u4e00\u79cd\u901a\u7528\u7f51\u7edc\u7ed3\u6784\uff0c\u52a0\u901f\u4e86 AI \u6a21\u578b\u7684\u5927\u4e00\u7edf\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u5b9e\u8df5\u8bc1\u660e\u5927\u6a21\u578b\u4e0d\u4ec5\u5728\u5e7f\u6cdb\u7684\u4efb\u52a1\u4e2d\u80fd\u4ea7\u751f\u66f4\u597d\u7684\u7ed3\u679c\u3001\u62e5\u6709\u66f4\u5f3a\u7684\u6cdb\u5316\u6027\uff0c\u8fd8\u53ef\u4ee5\u63d0\u5347\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u7387\uff0c\u751a\u81f3\u884d\u751f\u51fa\u65b0\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u5b66\u672f\u754c\u548c\u4ea7\u4e1a\u754c\u90fd\u5f00\u59cb\u8ffd\u6c42\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u7136\u800c\u968f\u7740\u6a21\u578b\u7684\u4e0d\u65ad\u6269\u5927\uff0c\u5176\u8bad\u7ec3\u8fc7\u7a0b\u4e5f\u53d8\u5f97\u66f4\u52a0\u56f0\u96be\uff0c\u6bd4\u5982\u4f1a\u51fa\u73b0\u8bad\u7ec3\u4e0d\u6536\u655b\u7b49\u95ee\u9898\u3002\u8fd9\u5c31\u9700\u8981\u5927\u91cf\u7684\u624b\u52a8\u8c03\u53c2\u5de5\u4f5c\u6765\u89e3\u51b3\uff0c\u800c\u8fd9\u4e0d\u4ec5\u4f1a\u9020\u6210\u8d44\u6e90\u6d6a\u8d39\uff0c\u8fd8\u4f1a\u4ea7\u751f\u4e0d\u53ef\u9884\u4f30\u7684\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n\n\n\n<p>\u4e0e\u5176\u626c\u6c64\u6b62\u6cb8\uff0c\u4e0d\u5982\u91dc\u5e95\u62bd\u85aa\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ece\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u7406\u8bba\u51fa\u53d1\uff0c\u521b\u65b0\u63a8\u51fa\u4e86 TorchScale \u5de5\u5177\u5305\uff0c\u5e76\u5df2\u5c06\u5176\u5f00\u6e90\u3002TorchScale \u662f\u4e00\u4e2a PyTorch \u5e93\uff0c\u5141\u8bb8\u79d1\u7814\u548c\u5f00\u53d1\u4eba\u5458\u66f4\u9ad8\u6548\u5730\u8bad\u7ec3 Transformer \u5927\u6a21\u578b\u3002\u540c\u65f6\uff0c\u5b83\u6709\u6548\u5730\u63d0\u5347\u4e86\u5efa\u6a21\u7684\u6027\u80fd\u548c\u901a\u7528\u6027\uff0c\u63d0\u9ad8\u4e86 Transformer \u7684\u7a33\u5b9a\u6027\u548c\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n\n\n\n<p>TorchScale GitHub \u9875\u9762\uff1ahttps:\/\/github.com\/microsoft\/torchscale<\/p>\n\n\n\n<p>\u201c\u6211\u4eec\u5e0c\u671b\u901a\u8fc7 TorchScale \u7684\u7cfb\u5217\u5de5\u4f5c\u4ece\u66f4\u5e95\u5c42\u51fa\u53d1\u505a\u4e00\u4e9b\u57fa\u7840\u6027\u7684\u7814\u7a76\u521b\u65b0\uff0c\u901a\u8fc7\u6570\u5b66\u6216\u8005\u7406\u8bba\u4e0a\u7684\u6307\u5bfc\u548c\u542f\u53d1\uff0c\u5728 Transformer \u6a21\u578b\u6269\u5c55\u7684\u5de5\u4f5c\u4e2d\u53d6\u5f97\u66f4\u597d\u7684\u6548\u679c\uff0c\u800c\u4e0d\u662f\u5355\u7eaf\u7684\u8c03\u53c2\u6216\u4ec5\u4ece\u5de5\u7a0b\u5c42\u9762\u53bb\u90e8\u5206\u7f13\u89e3\u67d0\u4e9b\u95ee\u9898\u3002TorchScale \u80fd\u591f\u652f\u6301\u4efb\u610f\u7684\u7f51\u7edc\u6df1\u5ea6\u548c\u5bbd\u5ea6\uff0c\u5b9e\u9a8c\u9a8c\u8bc1\u5b83\u53ef\u4ee5\u8f7b\u677e\u6269\u5927\u6a21\u578b\u89c4\u6a21\uff0c\u800c\u4e14\u53ea\u9700\u8981\u51e0\u884c\u4ee3\u7801\u5c31\u80fd\u591f\u5b9e\u73b0\u591a\u6a21\u6001\u6a21\u578b\u7684\u8bad\u7ec3\u3002\u201d\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u81ea\u7136\u8bed\u8a00\u8ba1\u7b97\u7ec4\u9996\u5e2d\u7814\u7a76\u5458\u97e6\u798f\u5982\u8868\u793a\u3002<\/p>\n\n\n\n<p>TorchScale \u4e3b\u8981\u4ece\u4e09\u4e2a\u65b9\u9762\u5e2e\u52a9\u79d1\u7814\u4eba\u5458\u514b\u670d\u4e86\u6269\u5c55 Transformer \u5927\u6a21\u578b\u65f6\u7684\u56f0\u96be\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>DeepNet\uff1a\u63d0\u5347\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li>Magneto\uff1a\u63d0\u5347\u6a21\u578b\u7684\u901a\u7528\u6027\u3002<\/li>\n\n\n\n<li>X-MoE\uff1a\u63d0\u5347\u6a21\u578b\u8bad\u7ec3\u7684\u9ad8\u6548\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"440\" height=\"466\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-1.png\" alt=\"\u56fe1\uff1aTorchScale \u89e3\u51b3\u4e86\u5927\u6a21\u578b\u5728\u7a33\u5b9a\u6027\u3001\u901a\u7528\u6027\u3001\u9ad8\u6548\u6027\u65b9\u9762\u7684\u95ee\u9898\" class=\"wp-image-1071069\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-1.png 440w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-1-283x300.png 283w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-1-170x180.png 170w\" sizes=\"auto, (max-width: 440px) 100vw, 440px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1aTorchScale \u89e3\u51b3\u4e86\u5927\u6a21\u578b\u5728\u7a33\u5b9a\u6027\u3001\u901a\u7528\u6027\u3001\u9ad8\u6548\u6027\u65b9\u9762\u7684\u95ee\u9898<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44088\"><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"deepnet-\u8ba9transformer\u8bad\u7ec3\u6df1\u5ea6\u8d85\u8fc71000\u5c42\">DeepNet\uff1a\u8ba9Transformer\u8bad\u7ec3\u6df1\u5ea6\u8d85\u8fc71000\u5c42<\/h2>\n\n\n\n<p>\u5c3d\u7ba1\u8fd1\u5e74\u6765\u6a21\u578b\u53c2\u6570\u7684\u6570\u91cf\u8d8a\u6765\u8d8a\u5927\uff0c\u5df2\u7ecf\u4ece\u767e\u4e07\u7ea7\u6269\u5c55\u5230\u4e07\u4ebf\u7ea7\uff0c\u4f46\u53c2\u6570\u7684\u6df1\u5ea6\u5374\u4e00\u76f4\u53d7\u9650\u4e8e Transformer \u8bad\u7ec3\u7684\u4e0d\u7a33\u5b9a\u6027\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u4e00\u4e9b\u79d1\u7814\u4eba\u5458\u5c1d\u8bd5\u901a\u8fc7\u66f4\u597d\u7684\u521d\u59cb\u5316\u6216\u67b6\u6784\u6765\u63d0\u5347 Transformer \u7684\u7a33\u5b9a\u6027\uff0c\u4f46\u8fd9\u4e5f\u53ea\u80fd\u8ba9 Transformer \u5728\u767e\u5c42\u7ea7\u522b\u7684\u6df1\u5ea6\u4e0b\u4fdd\u6301\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u6a21\u578b\u8f93\u51fa\u7684\u5267\u70c8\u53d8\u5316\u662f\u5bfc\u81f4\u6a21\u578b\u4e0d\u7a33\u5b9a\u7684\u91cd\u8981\u539f\u56e0\u3002\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u5728\u6b8b\u5dee\u8fde\u63a5\u5904\u4f7f\u7528\u4e86\u4e00\u79cd\u65b0\u7684\u5f52\u4e00\u5316\u51fd\u6570\u2014\u2014DeepNorm\u3002\u65b0\u7684\u51fd\u6570\u7531\u7406\u8bba\u63a8\u5bfc\u800c\u6765\uff0c\u53ef\u4ee5\u628a\u6a21\u578b\u8f93\u51fa\u7684\u53d8\u5316\u9650\u5236\u5728\u5e38\u6570\u8303\u56f4\u5185\u3002\u8fd9\u79cd\u65b9\u6cd5\u53ea\u9700\u8981\u6539\u53d8\u51e0\u884c\u4ee3\u7801\uff0c\u5c31\u53ef\u4ee5\u5927\u5e45\u63d0\u5347 Transformer \u7684\u7a33\u5b9a\u6027\u3002\u901a\u8fc7\u5f15\u5165\u65b0\u7684 DeepNorm \u51fd\u6570\uff0c\u7814\u7a76\u5458\u4eec\u8bad\u7ec3\u4e86\u8d85\u6df1\u7684 Transformer \u7f51\u7edc DeepNet\uff0c\u5728\u4fdd\u8bc1\u6a21\u578b\u7a33\u5b9a\u7684\u540c\u65f6\uff0c\u53ef\u4ee5\u5c06\u6a21\u578b\u6df1\u5ea6\u6269\u5c55\u52301000\u5c42\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<p>DeepNorm \u540c\u65f6\u5177\u5907 Post-LN \u7684\u6027\u80fd\u4f18\u52bf\u548c Pre-LN \u7684\u7a33\u5b9a\u8bad\u7ec3\u4f18\u52bf\u3002\u8fd9\u4e2a\u65b0\u65b9\u6cd5\u6216\u5c06\u6210\u4e3a Transformer \u7684\u9996\u9009\u66ff\u4ee3\u65b9\u6848\uff0c\u5b83\u4e0d\u4ec5\u9002\u7528\u4e8e\u6df1\u6a21\u578b\uff0c\u66f4\u9002\u7528\u4e8e\u5927\u6a21\u578b\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u4e0e\u5177\u6709120\u4ebf\u53c2\u6570\u768448\u5c42\u6a21\u578b\u76f8\u6bd4\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u966232\u4ebf\u53c2\u6570\u7684200\u5c42\u6a21\u578b\u5728100\u591a\u4e2a\u8bed\u8a00\u3001\u8d8510000\u4e2a\u8bed\u8a00\u5bf9\u548c130\u4ebf\u4e2a\u6587\u672c\u5bf9\u7684\u591a\u8bed\u8a00\u673a\u5668\u7ffb\u8bd1\u5b9e\u9a8c\u4e2d\u5b9e\u73b0\u4e865 BLEU \u7684\u63d0\u5347\u3002\u5728\u5927\u89c4\u6a21\u591a\u8bed\u8a00\u7ffb\u8bd1\u4efb\u52a1\u4e0a\uff0c\u968f\u7740 DeepNet \u6a21\u578b\u6df1\u5ea6\u4ece10\u5c42\u6269\u5c55\u81f3100\u5c42\u548c1000\u5c42\uff0c\u6a21\u578b\u4e5f\u83b7\u5f97\u4e86\u66f4\u9ad8\u7684 BLEU \u503c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"688\" height=\"470\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-2.png\" alt=\"\u56fe2\uff1a\u968f\u7740\u6a21\u578b\u6df1\u5ea6\u4ece10\u5c42\u6269\u5c55\u81f3100\u5c42\u548c1000\u5c42\uff0cDeepNet \u6709\u6548\u63d0\u5347\u4e86\u591a\u8bed\u8a00\u7ffb\u8bd1\u7ed3\u679c\" class=\"wp-image-1071072\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-2.png 688w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-2-300x205.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-2-240x164.png 240w\" sizes=\"auto, (max-width: 688px) 100vw, 688px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1a\u968f\u7740\u6a21\u578b\u6df1\u5ea6\u4ece10\u5c42\u6269\u5c55\u81f3100\u5c42\u548c1000\u5c42\uff0cDeepNet \u6709\u6548\u63d0\u5347\u4e86\u591a\u8bed\u8a00\u7ffb\u8bd1\u7ed3\u679c<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p id=\"caption-attachment-44089\"><\/p>\n\n\n\n<p>\u97e6\u798f\u5982\u8bf4\uff0c\u201c\u6b64\u524d\uff0c\u79d1\u7814\u4eba\u5458\u5728\u8bad\u7ec3\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u65f6\uff0c\u5f80\u5f80\u9700\u8981\u6295\u5165\u5927\u91cf\u7684\u7cbe\u529b\u5728\u6a21\u578b\u8c03\u53c2\u4e0a\uff0c\u65e0\u5f62\u4e2d\u589e\u52a0\u4e86\u5b9e\u9a8c\u6210\u672c\uff0c\u6709\u7684\u6a21\u578b\u5728\u8bad\u7ec3\u4e2d\u9014\u5c31\u65e0\u6cd5\u7ee7\u7eed\u4e0b\u53bb\u4e86\uff0c\u5373\u4f7f\u7ed9\u6a21\u578b\u6253\u4e0a\u8865\u4e01\u4e5f\u8fd8\u662f\u4f1a\u5f71\u54cd\u6a21\u578b\u6027\u80fd\u3002DeepNet \u53ef\u4ee5\u5e2e\u52a9\u79d1\u7814\u4eba\u5458\u5927\u5e45\u964d\u4f4e\u8c03\u53c2\u7684\u8d1f\u62c5\uff0c\u5728\u63d0\u5347\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\u964d\u4f4e\u5b9e\u9a8c\u6210\u672c\u3002\u201d<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"magneto-\u771f\u6b63\u5b9e\u73b0\u591a\u6a21\u6001\u6a21\u578b\u67b6\u6784\u7edf\u4e00\">Magneto\uff1a\u771f\u6b63\u5b9e\u73b0\u591a\u6a21\u6001\u6a21\u578b\u67b6\u6784\u7edf\u4e00<\/h2>\n\n\n\n<p>\u8de8\u8bed\u8a00\u3001\u89c6\u89c9\u3001\u8bed\u97f3\u548c\u591a\u6a21\u6001\u7684\u6a21\u578b\u5728\u6a21\u578b\u7ed3\u6784\u4e0a\u8d70\u5411\u5927\u4e00\u7edf\u7684\u8d8b\u52bf\u5982\u4eca\u5df2\u7ecf\u6108\u53d1\u660e\u663e\u3002\u5177\u4f53\u800c\u8a00\uff0c\u4ece NLP \u9886\u57df\u5f00\u59cb\uff0cTransformer \u5df2\u6210\u4e3a AI \u5404\u9886\u57df\u7684\u4e3b\u6d41\u7ed3\u6784\u3002\u7136\u800c\uff0c\u5c3d\u7ba1\u90fd\u4f7f\u7528\u4e86 Transformer\uff0c\u4f46\u4e0d\u540c\u6a21\u6001\u4efb\u52a1\u7684\u6a21\u578b\u7ed3\u6784\u5728\u5177\u4f53\u5b9e\u73b0\u65f6\u4ecd\u5b58\u5728\u663e\u8457\u5dee\u5f02\u3002\u4f8b\u5982\uff0cGPT \u548c ViT \u6a21\u578b\u91c7\u7528\u4e86 Pre-LN Transformer\uff0c\u800c BERT \u548c\u673a\u5668\u7ffb\u8bd1\u6a21\u578b\u4f7f\u7528\u7684\u662f Post-LN \u6765\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5bf9\u4e8e\u591a\u6a21\u6001\u6a21\u578b\uff0c\u4e0d\u540c\u8f93\u5165\u6a21\u6001\u7684\u6700\u4f18 Transformer \u53d8\u4f53\u901a\u5e38\u662f\u4e0d\u540c\u7684\u3002\u4ee5\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63a8\u51fa\u7684\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u6a21\u578b BEiT-3\u4e3a\u4f8b\uff0c\u5176\u4f7f\u7528 Post-LN \u5bf9\u4e8e\u89c6\u89c9\u90e8\u5206\u662f\u6b21\u4f18\u7684\uff0c\u800c Pre-LN \u5bf9\u4e8e\u8bed\u8a00\u90e8\u5206\u662f\u6b21\u4f18\u7684\u3002<\/p>\n\n\n\n<p>\u8981\u60f3\u8ba9\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u771f\u6b63\u5b9e\u73b0\u5927\u4e00\u7edf\u5c31\u9700\u8981\u4e00\u4e2a\u7edf\u4e00\u7684\u67b6\u6784\uff0c\u8be5\u67b6\u6784\u9700\u8981\u5728\u4e0d\u540c\u4efb\u52a1\u548c\u6a21\u6001\u4e0a\u90fd\u80fd\u6709\u826f\u597d\u7684\u6027\u80fd\u8868\u73b0\u3002\u53e6\u5916\uff0c\u5982\u4e4b\u524d\u6240\u8ff0\uff0cTransformer \u67b6\u6784\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u4e5f\u662f\u4e00\u4e2a\u75db\u70b9\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u610f\u8bc6\u5230\uff0c\u901a\u7528\u6a21\u578b\u7684\u5f00\u53d1\u9700\u8981\u66f4\u57fa\u7840\u7684 Transformer\uff0c\u5373 Foundation Transformer\u3002\u9996\u5148\u5b83\u7684\u5efa\u6a21\u80fd\u591f\u4f5c\u4e3a\u5404\u79cd\u4efb\u52a1\u548c\u6a21\u5f0f\u7684\u7edf\u4e00\u67b6\u6784\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u4f7f\u7528\u76f8\u540c\u7684\u4e3b\u5e72\u800c\u65e0\u9700\u53cd\u590d\u9b54\u6539\u3002\u5176\u901a\u7528\u7684\u8bbe\u8ba1\u539f\u5219\u4e5f\u5e94\u8be5\u652f\u6301\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\u7684\u5f00\u53d1\uff0c\u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u5c06\u7edf\u4e00\u7684 Transformer \u7528\u4e8e\u5404\u79cd\u6a21\u6001\u3002\u5176\u6b21\uff0c\u5b83\u7684\u7f51\u7edc\u7ed3\u6784\u5e94\u80fd\u591f\u4fdd\u969c\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\uff0c\u4ece\u800c\u964d\u4f4e\u57fa\u7840\u6a21\u578b\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u96be\u5ea6\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e9b\u76ee\u6807\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a Foundation Transformer\u2014\u2014Magneto\u3002\u5728 Magneto \u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u5f15\u5165\u4e86 Sub-LN\uff0c\u4e3a\u6bcf\u4e2a\u5b50\u5c42\uff08\u5373\u591a\u5934\u81ea\u6ce8\u610f\u529b\u548c\u524d\u9988\u7f51\u7edc\uff09\u6dfb\u52a0\u4e86\u989d\u5916\u7684 LayerNorm\uff0c\u5e76\u4e14\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u521d\u59cb\u5316\u65b9\u6cd5\uff0c\u4e3a\u4ece\u6839\u672c\u4e0a\u63d0\u9ad8\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u63d0\u4f9b\u4e86\u7406\u8bba\u4fdd\u8bc1\u3002<\/p>\n\n\n\n<p>\u901a\u8fc7\u5bf9 Magneto \u5728\u4e0d\u540c\u4efb\u52a1\u548c\u6a21\u6001\u4e0a\u7684\u8bc4\u6d4b\uff0c\u5305\u62ec\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff08\u5373BERT\uff09\u3001\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\uff08\u4f8b\u5982GPT\uff09\u3001\u673a\u5668\u7ffb\u8bd1\u3001\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08\u5373BEiT\uff09\u3001\u8bed\u97f3\u8bc6\u522b\u548c\u89c6\u89c9\u8bed\u8a00\u9884\u8bad\u7ec3\uff08\u5373BEiT-3\uff09\uff0c\u7ed3\u679c\u663e\u793a\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\uff0cMagneto \u663e\u8457\u4f18\u4e8e\u5404\u79cd Transformer \u53d8\u4f53\u3002\u6b64\u5916\uff0c\u5f97\u76ca\u4e8e\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u63d0\u9ad8\uff0cMagneto \u8fd8\u5141\u8bb8\u4f7f\u7528\u66f4\u9ad8\u7684\u5b66\u4e60\u7387\u6765\u8fdb\u4e00\u6b65\u63d0\u9ad8\u7ed3\u679c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"472\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3-1024x472.jpg\" alt=\"\u56fe3\uff1aMagneto \u5728\u8bed\u8a00\u3001\u56fe\u50cf\u3001\u8bed\u97f3\u548c\u591a\u6a21\u6001\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\" class=\"wp-image-1071075\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3-1024x472.jpg 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3-300x138.jpg 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3-768x354.jpg 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3-240x111.jpg 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-3.jpg 1270w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1aMagneto \u5728\u8bed\u8a00\u3001\u56fe\u50cf\u3001\u8bed\u97f3\u548c\u591a\u6a21\u6001\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p id=\"caption-attachment-44090\"><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"x-moe-\u4f18\u4e8e\u57fa\u7ebfsmoe\u6a21\u578b-\u52a9\u529b\u6a21\u578b\u9ad8\u6548\u8bad\u7ec3\">X-MoE\uff1a\u4f18\u4e8e\u57fa\u7ebfSMoE\u6a21\u578b\uff0c\u52a9\u529b\u6a21\u578b\u9ad8\u6548\u8bad\u7ec3<\/h2>\n\n\n\n<p>\u5728\u6709\u5173\u5927\u6a21\u578b\u8bad\u7ec3\u7684\u7814\u7a76\u4e2d\uff0c\u9664\u4e86\u5c06\u7f51\u7edc\u6df1\u5ea6\u505a\u5f97\u66f4\u6df1\u548c\u5c06\u5bbd\u5ea6\u5373\u9690\u85cf\u7ef4\u5ea6\u6269\u5927\u4ee5\u5916\uff0c\u8fd8\u53ef\u4ee5\u5229\u7528\u6df7\u5408\u4e13\u5bb6\u7cfb\u7edf\uff08Mixture of Experts, MoE\uff09\u3002\u5c3d\u7ba1 MoE \u53ef\u4ee5\u5728\u8bf8\u5982\u8bed\u8a00\u6a21\u578b\u548c\u89c6\u89c9\u8868\u793a\u5b66\u4e60\u7b49\u5e7f\u6cdb\u95ee\u9898\u4e0a\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\uff0c\u4f46\u4e5f\u4f1a\u5bfc\u81f4\u66f4\u9ad8\u7684\u8ba1\u7b97\u6210\u672c\uff0c\u8fd9\u4fc3\u4f7f\u8d8a\u6765\u8d8a\u591a\u7684\u79d1\u7814\u4eba\u5458\u5f00\u59cb\u63a2\u7d22\u7a00\u758f\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Sparse Mixture-of-Experts, SMoE\uff09\u3002SMoE \u4e3b\u8981\u901a\u8fc7\u6784\u5efa\u7a00\u758f\u6fc0\u6d3b\u7684\u795e\u7ecf\u7f51\u7edc\u6765\u589e\u52a0\u6a21\u578b\u5bb9\u91cf\u3002\u5728\u4e0d\u663e\u8457\u589e\u52a0\u8ba1\u7b97\u5f00\u9500\u7684\u60c5\u51b5\u4e0b SMoE \u6a21\u578b\u5728\u5404\u79cd\u4efb\u52a1\uff08\u5305\u62ec\u673a\u5668\u7ffb\u8bd1\u3001\u56fe\u50cf\u5206\u7c7b\u548c\u8bed\u97f3\u8bc6\u522b\uff09\u4e0a\u7684\u6027\u80fd\u90fd\u4f18\u4e8e\u7a20\u5bc6\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u5728 SMoE \u6a21\u578b\u4e2d\uff0c\u8def\u7531\u673a\u5236\u53d1\u6325\u7740\u91cd\u8981\u7684\u4f5c\u7528\u3002\u7ed9\u5b9a\u8f93\u5165 token\uff0c\u8def\u7531\u673a\u5236\u4f1a\u6d4b\u91cf\u6bcf\u4e2a token \u4e0e\u4e13\u5bb6\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u5206\u6570\uff0c\u7136\u540e\u518d\u6839\u636e\u8def\u7531\u5f97\u5206\u5c06 token \u5206\u914d\u7ed9\u6700\u5339\u914d\u7684\u4e13\u5bb6\u3002\u56e0\u6b64\uff0c\u8fd1\u5e74\u6765\u8bb8\u591a\u7814\u7a76\u90fd\u96c6\u4e2d\u5728\u5982\u4f55\u8bbe\u8ba1 token \u4e13\u5bb6\u5206\u914d\u7b97\u6cd5\u4e0a\u3002\u7136\u800c\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u5f53\u524d\u7684\u8def\u7531\u673a\u5236\u503e\u5411\u4e8e\u4ee5\u4e13\u5bb6\u4e3a\u4e2d\u5fc3\u6765\u63a8\u52a8\u9690\u85cf\u8868\u793a\u805a\u7c7b\uff0c\u8fd9\u5bb9\u6613\u5f15\u8d77\u8868\u5f81\u574d\u584c\uff08Representation Collapse\uff09\uff0c\u635f\u5bb3\u6a21\u578b\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u7f13\u89e3\u73b0\u6709\u7684\u8def\u7531\u673a\u5236\u5f15\u8d77\u7684\u8868\u5f81\u574d\u584c\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u65b0\u7684\u65b9\u6cd5 X-MoE\uff0c\u4e3a SMoE \u6a21\u578b\u5f15\u5165\u4e86\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u8def\u7531\u7b97\u6cd5\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u533a\u522b\u4e8e\u73b0\u6709 SMoE \u6a21\u578b\u76f4\u63a5\u4f7f\u7528\u9690\u85cf\u5411\u91cf\u8fdb\u884c\u8def\u7531\uff0cX-MoE \u5148\u5c06\u9690\u85cf\u5411\u91cf\u6295\u5c04\u5230\u4f4e\u7ef4\u7a7a\u95f4\u4e2d\uff0c\u518d\u5bf9 token \u8868\u793a\u548c\u4e13\u5bb6\u8868\u793a\u8fdb\u884c L2 \u5f52\u4e00\u5316\uff0c\u6765\u6d4b\u91cf\u4f4e\u7ef4\u8d85\u7403\u9762\u4e0a\u7684\u8def\u7531\u5206\u6570\u3002\u6b64\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u63d0\u51fa\u4e86\u8f6f\u4e13\u5bb6\u95e8\uff08soft expert gate\uff09\uff0c\u4ee5\u5b66\u4e60\u63a7\u5236\u4e13\u5bb6\u7684\u6fc0\u6d3b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"501\" height=\"358\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-4.png\" alt=\"\u56fe4\uff1aX-MoE \u6d41\u7a0b\u56fe\" class=\"wp-image-1071078\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-4.png 501w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-4-300x214.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/torchscale-4-240x171.png 240w\" sizes=\"auto, (max-width: 501px) 100vw, 501px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1aX-MoE \u6d41\u7a0b\u56fe<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p id=\"caption-attachment-44091\"><\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5bf9\u8fd9\u4e00\u65b0\u65b9\u6cd5\u5728\u8de8\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u8bc4\u6d4b\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u8bed\u8a00\u5efa\u6a21\u548c\u5fae\u8c03\u6027\u80fd\u65b9\u9762\uff0c\u57fa\u4e8e X-MoE \u7684\u6a21\u578b\u59cb\u7ec8\u4f18\u4e8e\u57fa\u7ebf SMoE \u6a21\u578b\u3002\u5b9e\u9a8c\u5206\u6790\u8fd8\u8868\u660e\uff0c\u4e0e SMoE \u57fa\u7ebf\u76f8\u6bd4\uff0cX-MoE \u65b9\u6cd5\u6709\u6548\u7f13\u89e3\u4e86\u8868\u5f81\u574d\u584c\u95ee\u9898\u3002\u8be5\u65b9\u6cd5\u5728\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u671f\u95f4\u4e5f\u5b9e\u73b0\u4e86\u66f4\u4e00\u81f4\u7684\u8def\u7531\u884c\u4e3a\uff0c\u8bc1\u5b9e\u4e86 X-MoE \u8def\u7531\u7b97\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u201c\u968f\u7740\u6280\u672f\u7684\u6301\u7eed\u6f14\u8fdb\uff0c\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u4e0d\u4ec5\u4ec5\u662f\u5de5\u7a0b\u5c42\u9762\u7684\u5de5\u4f5c\u3002\u6211\u4eec\u5e94\u8be5\u4ece\u57fa\u7840\u7814\u7a76\u7684\u89d2\u5ea6\u51fa\u53d1\uff0c\u63a2\u7d22\u4e0b\u4e00\u4ee3 Transformer \u7f51\u7edc\u67b6\u6784\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5728 AI \u6a21\u578b\u5927\u4e00\u7edf\u8d8b\u52bf\u7684\u63a8\u52a8\u4e0b\uff0c\u6211\u4eec\u66f4\u5e94\u8be5\u8ffd\u6c42\u540c\u4e00\u7ed3\u6784\u6765\u652f\u6301\u4e0d\u540c\u6a21\u6001\u7684\u8f93\u5165\uff0c\u5e76\u5728\u4e0d\u540c\u8bed\u8a00\u548c\u6a21\u6001\u7684\u4efb\u52a1\u4e0a\u83b7\u5f97\u826f\u597d\u7684\u6027\u80fd\u3002\u901a\u8fc7\u7406\u8bba\u6307\u5bfc\u8ba9\u6a21\u578b\u53d8\u5f97\u66f4\u5927\u3001\u66f4\u7a33\u5b9a\u3001\u66f4\u901a\u7528\u3002\u201d\u97e6\u798f\u5982\u8bf4\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7f16\u8005\u6309\uff1a\u8fd1\u671f\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ece\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u7406\u8bba\u51fa\u53d1\uff0c\u7814\u53d1\u5e76\u63a8\u51fa\u4e86 TorchScale \u5f00\u6e90\u5de5\u5177\u5305\u3002TorchScale \u5de5\u5177\u5305\u901a\u8fc7\u91c7\u7528 DeepNet\u3001Magneto \u548c X-MoE \u7b49\u6700\u5148\u8fdb\u7684\u5efa\u6a21\u6280\u672f\uff0c\u53ef\u4ee5\u5e2e\u52a9\u7814\u7a76\u548c\u5f00\u53d1\u4eba\u5458\u63d0\u9ad8\u5efa\u6a21\u7684\u901a\u7528\u6027\u548c\u6574\u4f53\u6027\u80fd\uff0c\u786e\u4fdd\u8bad\u7ec3\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u53ca\u6548\u7387\uff0c\u5e76\u5141\u8bb8\u4ee5\u4e0d\u540c\u7684\u6a21\u578b\u5927\u5c0f\u6269\u5c55 Transformer \u7f51\u7edc\u3002 \u5982\u4eca\uff0c\u5728\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\uff08CV\uff09\u3001\u8bed\u97f3\u3001\u591a\u6a21\u6001\u6a21\u578b\u548c AI for Science \u7b49\u9886\u57df\u7814\u7a76\u4e2d\uff0cTransformer \u5df2\u7ecf\u6210\u4e3a\u4e00\u79cd\u901a\u7528\u7f51\u7edc\u7ed3\u6784\uff0c\u52a0\u901f\u4e86 AI \u6a21\u578b\u7684\u5927\u4e00\u7edf\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u5b9e\u8df5\u8bc1\u660e\u5927\u6a21\u578b\u4e0d\u4ec5\u5728\u5e7f\u6cdb\u7684\u4efb\u52a1\u4e2d\u80fd\u4ea7\u751f\u66f4\u597d\u7684\u7ed3\u679c\u3001\u62e5\u6709\u66f4\u5f3a\u7684\u6cdb\u5316\u6027\uff0c\u8fd8\u53ef\u4ee5\u63d0\u5347\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u7387\uff0c\u751a\u81f3\u884d\u751f\u51fa\u65b0\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u5b66\u672f\u754c\u548c\u4ea7\u4e1a\u754c\u90fd\u5f00\u59cb\u8ffd\u6c42\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002 \u7136\u800c\u968f\u7740\u6a21\u578b\u7684\u4e0d\u65ad\u6269\u5927\uff0c\u5176\u8bad\u7ec3\u8fc7\u7a0b\u4e5f\u53d8\u5f97\u66f4\u52a0\u56f0\u96be\uff0c\u6bd4\u5982\u4f1a\u51fa\u73b0\u8bad\u7ec3\u4e0d\u6536\u655b\u7b49\u95ee\u9898\u3002\u8fd9\u5c31\u9700\u8981\u5927\u91cf\u7684\u624b\u52a8\u8c03\u53c2\u5de5\u4f5c\u6765\u89e3\u51b3\uff0c\u800c\u8fd9\u4e0d\u4ec5\u4f1a\u9020\u6210\u8d44\u6e90\u6d6a\u8d39\uff0c\u8fd8\u4f1a\u4ea7\u751f\u4e0d\u53ef\u9884\u4f30\u7684\u8ba1\u7b97\u6210\u672c\u3002 \u4e0e\u5176\u626c\u6c64\u6b62\u6cb8\uff0c\u4e0d\u5982\u91dc\u5e95\u62bd\u85aa\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ece\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u7406\u8bba\u51fa\u53d1\uff0c\u521b\u65b0\u63a8\u51fa\u4e86 TorchScale \u5de5\u5177\u5305\uff0c\u5e76\u5df2\u5c06\u5176\u5f00\u6e90\u3002TorchScale \u662f\u4e00\u4e2a PyTorch \u5e93\uff0c\u5141\u8bb8\u79d1\u7814\u548c\u5f00\u53d1\u4eba\u5458\u66f4\u9ad8\u6548\u5730\u8bad\u7ec3 Transformer \u5927\u6a21\u578b\u3002\u540c\u65f6\uff0c\u5b83\u6709\u6548\u5730\u63d0\u5347\u4e86\u5efa\u6a21\u7684\u6027\u80fd\u548c\u901a\u7528\u6027\uff0c\u63d0\u9ad8\u4e86 Transformer \u7684\u7a33\u5b9a\u6027\u548c\u8bad\u7ec3\u6548\u7387\u3002 TorchScale GitHub \u9875\u9762\uff1ahttps:\/\/github.com\/microsoft\/torchscale \u201c\u6211\u4eec\u5e0c\u671b\u901a\u8fc7 TorchScale \u7684\u7cfb\u5217\u5de5\u4f5c\u4ece\u66f4\u5e95\u5c42\u51fa\u53d1\u505a\u4e00\u4e9b\u57fa\u7840\u6027\u7684\u7814\u7a76\u521b\u65b0\uff0c\u901a\u8fc7\u6570\u5b66\u6216\u8005\u7406\u8bba\u4e0a\u7684\u6307\u5bfc\u548c\u542f\u53d1\uff0c\u5728 Transformer \u6a21\u578b\u6269\u5c55\u7684\u5de5\u4f5c\u4e2d\u53d6\u5f97\u66f4\u597d\u7684\u6548\u679c\uff0c\u800c\u4e0d\u662f\u5355\u7eaf\u7684\u8c03\u53c2\u6216\u4ec5\u4ece\u5de5\u7a0b\u5c42\u9762\u53bb\u90e8\u5206\u7f13\u89e3\u67d0\u4e9b\u95ee\u9898\u3002TorchScale \u80fd\u591f\u652f\u6301\u4efb\u610f\u7684\u7f51\u7edc\u6df1\u5ea6\u548c\u5bbd\u5ea6\uff0c\u5b9e\u9a8c\u9a8c\u8bc1\u5b83\u53ef\u4ee5\u8f7b\u677e\u6269\u5927\u6a21\u578b\u89c4\u6a21\uff0c\u800c\u4e14\u53ea\u9700\u8981\u51e0\u884c\u4ee3\u7801\u5c31\u80fd\u591f\u5b9e\u73b0\u591a\u6a21\u6001\u6a21\u578b\u7684\u8bad\u7ec3\u3002\u201d\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u81ea\u7136\u8bed\u8a00\u8ba1\u7b97\u7ec4\u9996\u5e2d\u7814\u7a76\u5458\u97e6\u798f\u5982\u8868\u793a\u3002 TorchScale \u4e3b\u8981\u4ece\u4e09\u4e2a\u65b9\u9762\u5e2e\u52a9\u79d1\u7814\u4eba\u5458\u514b\u670d\u4e86\u6269\u5c55 Transformer \u5927\u6a21\u578b\u65f6\u7684\u56f0\u96be\uff1a \u5c3d\u7ba1\u8fd1\u5e74\u6765\u6a21\u578b\u53c2\u6570\u7684\u6570\u91cf\u8d8a\u6765\u8d8a\u5927\uff0c\u5df2\u7ecf\u4ece\u767e\u4e07\u7ea7\u6269\u5c55\u5230\u4e07\u4ebf\u7ea7\uff0c\u4f46\u53c2\u6570\u7684\u6df1\u5ea6\u5374\u4e00\u76f4\u53d7\u9650\u4e8e Transformer \u8bad\u7ec3\u7684\u4e0d\u7a33\u5b9a\u6027\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u4e00\u4e9b\u79d1\u7814\u4eba\u5458\u5c1d\u8bd5\u901a\u8fc7\u66f4\u597d\u7684\u521d\u59cb\u5316\u6216\u67b6\u6784\u6765\u63d0\u5347 Transformer \u7684\u7a33\u5b9a\u6027\uff0c\u4f46\u8fd9\u4e5f\u53ea\u80fd\u8ba9 Transformer \u5728\u767e\u5c42\u7ea7\u522b\u7684\u6df1\u5ea6\u4e0b\u4fdd\u6301\u7a33\u5b9a\u3002 \u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u6a21\u578b\u8f93\u51fa\u7684\u5267\u70c8\u53d8\u5316\u662f\u5bfc\u81f4\u6a21\u578b\u4e0d\u7a33\u5b9a\u7684\u91cd\u8981\u539f\u56e0\u3002\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u5728\u6b8b\u5dee\u8fde\u63a5\u5904\u4f7f\u7528\u4e86\u4e00\u79cd\u65b0\u7684\u5f52\u4e00\u5316\u51fd\u6570\u2014\u2014DeepNorm\u3002\u65b0\u7684\u51fd\u6570\u7531\u7406\u8bba\u63a8\u5bfc\u800c\u6765\uff0c\u53ef\u4ee5\u628a\u6a21\u578b\u8f93\u51fa\u7684\u53d8\u5316\u9650\u5236\u5728\u5e38\u6570\u8303\u56f4\u5185\u3002\u8fd9\u79cd\u65b9\u6cd5\u53ea\u9700\u8981\u6539\u53d8\u51e0\u884c\u4ee3\u7801\uff0c\u5c31\u53ef\u4ee5\u5927\u5e45\u63d0\u5347 Transformer \u7684\u7a33\u5b9a\u6027\u3002\u901a\u8fc7\u5f15\u5165\u65b0\u7684 DeepNorm \u51fd\u6570\uff0c\u7814\u7a76\u5458\u4eec\u8bad\u7ec3\u4e86\u8d85\u6df1\u7684 Transformer [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1071081,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1071024","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071024","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":2,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071024\/revisions"}],"predecessor-version":[{"id":1071084,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071024\/revisions\/1071084"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1071081"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1071024"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1071024"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1071024"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1071024"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}