{"id":36209,"date":"2025-01-07T15:00:12","date_gmt":"2025-01-07T07:00:12","guid":{"rendered":"https:\/\/17aitech.com\/?p=36209"},"modified":"2025-01-07T15:00:12","modified_gmt":"2025-01-07T07:00:12","slug":"jurgen%e3%80%81%e6%9b%bc%e5%ae%81%e7%ad%89%e5%a4%a7%e4%bd%ac%e6%96%b0%e4%bd%9c%ef%bc%9amoe%e9%87%8d%e5%a1%916%e5%b9%b4%e5%89%8d%e7%9a%84universal-transformer%ef%bc%8c%e9%ab%98%e6%95%88%e5%8d%87","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36209","title":{"rendered":"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-19-2\" target=\"_blank\">Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7<\/a><\/p>\n<section>7 \u5e74\u524d\uff0c\u8c37\u6b4c\u5728\u8bba\u6587\u300aAttention is All You Need\u300b\u4e2d\u63d0\u51fa\u4e86 Transformer\u3002\u5c31\u5728 Transformer \u63d0\u51fa\u7684\u7b2c\u4e8c\u5e74\uff0c\u8c37\u6b4c\u53c8\u53d1\u5e03\u4e86\u00a0Universal Transformer\uff08UT\uff09\u3002\u5b83\u7684\u6838\u5fc3\u7279\u5f81\u662f\u901a\u8fc7\u8de8\u5c42\u5171\u4eab<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6765\u5b9e\u73b0\u6df1\u5ea6\u5faa\u73af\uff0c\u4ece\u800c\u91cd\u65b0\u5f15\u5165\u4e86 RNN \u5177\u6709\u7684\u5faa\u73af\u8868\u8fbe\u80fd\u529b\u3002\u5c42\u5171\u4eab\u4f7f\u5f97 UT \u5728<mark data-type=\"concepts\" data-id=\"95a97f4b-79d2-4bbc-91ae-300f074dff9f\">\u903b\u8f91<\/mark>\u63a8\u7406\u4efb\u52a1\u7b49\u7ec4\u5408\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u4f18\u4e8e Transformer\uff0c\u540c\u65f6\u8fd8\u5728\u5c0f\u89c4\u6a21\u8bed\u8a00\u5efa\u6a21\u548c\u7ffb\u8bd1\u4efb\u52a1\u4e0a\u5f97\u5230\u6539\u8fdb\u3002<\/section>\n<section><\/section>\n<section>UT \u5df2\u88ab\u8bc1\u660e\u5177\u6709\u66f4\u597d\u7684\u7ec4\u5408\u6cdb\u5316\u7279\u6027\uff0c\u80fd\u591f\u5728\u65e0\u76d1\u7763\u7684\u60c5\u51b5\u4e0b\u89e3\u6784\u7ed3\u6784\u5316\u95ee\u9898\u5e76\u6cdb\u5316\u5230\u66f4\u957f\u7684\u5e8f\u5217\u3002\u56e0\u6b64\u4e0e Transformer \u76f8\u6bd4\uff0cUT \u662f\u4e00\u79cd\u5177\u6709\u5353\u8d8a\u6cdb\u5316\u7279\u6027\u7684\u901a\u7528\u6027\u66f4\u5f3a\u7684\u67b6\u6784\u3002<\/section>\n<section><\/section>\n<section>\u4f46 UT \u7684\u8ba1\u7b97\u6548\u7387\u8fdc\u4f4e\u4e8e\u6807\u51c6 Transformer\uff0c\u4e0d\u9002\u5408\u5f53\u524d\u8bed\u8a00\u5efa\u6a21\u7b49\u4ee5<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u4e3a\u738b\u7684\u4efb\u52a1\u3002\u90a3\u4e48\uff0c\u6211\u4eec\u80fd\u4e0d\u80fd\u5f00\u53d1\u51fa\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8\u7684 UT \u6a21\u578b\uff0c\u5e76\u8fd9\u7c7b\u4efb\u52a1\u4e0a\u5b9e\u73b0\u6bd4\u6807\u51c6 Transformer \u66f4\u5177\u7ade\u4e89\u529b\u7684\u6027\u80fd\u5462\uff1f<\/section>\n<section><\/section>\n<section>\u8fd1\u65e5\uff0c\u5305\u62ec LSTM \u4e4b\u7236 J\u00fcrgen Schmidhuber\u3001\u65af\u5766\u798f\u5927\u5b66\u6559\u6388 Christopher Manning \u7b49\u5728\u5185\u7684\u7814\u7a76\u8005\u4ece\u5168\u65b0\u89c6\u89d2\u51fa\u53d1\uff0c\u63d0\u51fa\u4e86\u89e3\u51b3 UT \u57fa\u7840\u8ba1\u7b97<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6bd4\u95ee\u9898\u7684\u6700\u65b0\u65b9\u6848\u3002\u5177\u4f53\u6765\u8bb2\uff0c<strong>\u4ed6\u4eec\u63d0\u51fa Mixture-of-Experts Universal Transformers\uff08\u7b80\u79f0 MoEUT\uff09\uff0c\u5b83\u662f\u4e00\u79cd\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u67b6\u6784\uff0c\u5141\u8bb8 UT \u4ee5\u8ba1\u7b97\u548c\u5185\u5b58\u9ad8\u6548\u7684\u65b9\u5f0f\u6269\u5c55\u3002<\/strong><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aMoEUT: Mixture-of-Experts Universal Transformers<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2405.16039<\/section>\n<\/li>\n<li>\n<section>\u4ee3\u7801\u5730\u5740\uff1ahttps:\/\/github.com\/robertcsordas\/moeut<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5728\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u5229\u7528\u4e86 MoE \u5728\u524d\u9988\u548c<mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\u5c42\u65b9\u9762\u7684\u5404\u79cd\u6700\u65b0\u8fdb\u5c55\uff0c\u5e76\u5c06\u8fd9\u4e9b\u8fdb\u5c55\u4e0e\u4ee5\u4e0b\u4e24\u9879\u521b\u65b0\u5de5\u4f5c\u76f8\u7ed3\u5408\uff1a1\uff09layer grouping\uff0c\u5faa\u73af<mark data-type=\"tech_methods\" data-id=\"efb128f3-0b7c-45a8-a6ae-065fa93c6556\">\u5806\u53e0<\/mark> MoE \u5c42\u7ec4\uff1b2\uff09peri-layernorm \u65b9\u6848\uff08\u4f4d\u4e8e pre-layernorm \u548c post-layernorm \u4e4b\u95f4\uff09\uff0c\u5e76\u4e14\u4ec5\u5728\u7d27\u63a5 sigmoid \u6216 softmax \u6fc0\u6d3b\u4e4b\u524d\u5e94\u7528\u5c42<mark data-type=\"concepts\" data-id=\"0caf290b-4ff5-445b-8f62-ecc75a6a237f\">\u8303\u6570<\/mark>\u3002\u8fd9\u4e24\u8005\u90fd\u662f\u4e13\u4e3a\u5171\u4eab\u5c42 MoE \u67b6\u6784\u8bbe\u8ba1\uff0c\u5e76\u4e14\u6709\u5f3a\u6709\u529b\u7684\u5b9e\u8bc1\u8bc1\u636e\u652f\u6301\u3002<\/section>\n<section><\/section>\n<section>\u4ece\u5176\u4f5c\u7528\u6765\u8bb2\uff0cMoEUT \u5141\u8bb8\u6784\u5efa<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u548c\u8d44\u6e90\u9ad8\u6548\u7684 UT <mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff0c\u4e0d\u4ec5\u5728\u6211\u4eec\u53ef\u4ee5\u8d1f\u62c5\u5f97\u8d77\u7684\u6240\u6709\u89c4\u6a21\uff08\u6700\u9ad8 10 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\uff09\u4e0a\u5bf9\u7b97\u529b\u548c\u5185\u5b58\u7684\u9700\u6c42\u66f4\u4f4e\uff0c\u6027\u80fd\u4e5f\u8d85\u8d8a\u4e86\u6807\u51c6 Transformer\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u5728 C4\u3001SlimPajama \u548c peS2o \u8bed\u8a00\u5efa\u6a21\u6570\u636e\u96c6\u3001\u4ee5\u53ca The Stack \u4ee3\u7801\u6570\u636e\u96c6\u4e0a\u6d4b\u8bd5 MoEUT \u7684\u80fd\u529b\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u5faa\u73af\u5bf9\u4e8e\u6a21\u578b\u5b9e\u73b0\u5177\u6709\u7ade\u4e89\u529b\u7684\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002\u540c\u6837\u5730\uff0c\u7814\u7a76\u8005\u5728 BLiMP \u548c\u513f\u7ae5\u56fe\u4e66\u6d4b\u8bd5\u3001Lambada\u3001HellaSwag\u3001PIQA \u548c ARC-E \u7b49\u4e0b\u6e38\u4efb\u52a1\u4e0a\u5c55\u73b0\u4e86\u826f\u597d\u7684\u96f6\u6837\u672c\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>MoEUT \u67b6\u6784\u6982\u89c8<\/strong><\/section>\n<section><\/section>\n<section>\u5982\u524d\u6587\u6240\u8ff0\uff0cMoEUT \u662f\u4e00\u79cd\u5177\u6709\u5c42\u5171\u4eab<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684 Transformer \u67b6\u6784\uff0c\u5176\u4e2d\u4f7f\u7528 MoE \u6765\u89e3\u51b3<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u8ba1\u7b97\u6bd4\u95ee\u9898\u3002\u867d\u7136\u6700\u8fd1\u51fa\u73b0\u4e86\u5f88\u591a\u5173\u4e8e Transformer <mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684 MoE \u65b9\u6cd5\u7814\u7a76\uff0c\u4f46\u8981\u8ba9\u5b83\u4eec\u5728<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u76f8\u540c\u7684\u60c5\u51b5\u4e0b\u4e0e\u5bc6\u96c6\u65b9\u6cd5\u7ade\u4e89\uff0c\u4ecd\u7136\u662f\u4e00\u9879\u8270\u5de8\u7684\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u7814\u7a76\u8005\u5229\u7528\u4e86 MoE \u524d\u9988\u7f51\u7edc\u5757\uff08FFN\uff09\u3001MoE <mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\u5c42\uff0c\u5e76\u5f15\u5165\u4e86\u4e24\u79cd\u8003\u8651\u5230\u5171\u4eab\u5c42\u6a21\u578b\u7279\u5b9a\u5c5e\u6027\u7684\u65b0\u65b9\u6cd5 \u2014\u2014layer grouping \u548c\u4fe1\u53f7\u4f20\u64ad\u3002\u8fd9\u4e9b\u6280\u672f\u7684\u7ed3\u5408\u5bf9\u4e8e\u5b9e\u73b0\u6709\u6548\u7684\u5171\u4eab\u5c42 MoE Transformer \u53d1\u6325\u4e86\u5de8\u5927\u4f5c\u7528\u3002<\/section>\n<section><\/section>\n<section><strong>MoE \u524d\u9988\u5757<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u901a\u8fc7 MoE \u6765<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5316\u5171\u4eab\u5c42 Transformer \u7684\u524d\u9988\u5757\uff0c\u7814\u7a76\u8005\u4f7f\u7528\u4e86 \u03c3-MoE \u5e76\u505a\u4e86\u4e00\u4e9b\u4fee\u6539\u3002\u03c3-MoE \u5c06\u524d\u9988\u5757\u5212\u5206\u4e3a N_E \u4e2a\u5207\u7247\uff0c\u79f0\u4e3a\u4e13\u5bb6\uff08expert\uff09\u3002\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u6709\u4e24\u7ec4<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\uff0c\u5206\u522b\u662f\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-08cc7bd4836895909fd49f9ebecc3726.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-08cc7bd4836895909fd49f9ebecc3726.png\"><\/a>\u00a0\u548c\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9a9530ec59fd2a716554561cd944dfbb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9a9530ec59fd2a716554561cd944dfbb.png\"><\/a>\uff0c\u5176\u4e2d e \u2208{1,&#8230;,N_E} \u662f\u4e13\u5bb6\u7d22\u5f15\u3002<\/section>\n<section><\/section>\n<section>\u5728\u6bcf\u4e2a token \u4f4d\u7f6e t\uff0c\u7ed9\u5b9a\u5c42\u8f93\u5165 x_t \u2208 R^d_model\uff0cMoE \u524d\u9988\u5c42\u4f1a\u4e3a\u6bcf\u4e2a\u4e13\u5bb6\u8ba1\u7b97\u4e00\u4e2a\u5206\u6570\uff0c\u4ece\u800c\u5f97\u5230\u4e00\u4e2a\u5411\u91cf s \u2208 R^N_E\uff0c\u5176\u8ba1\u7b97\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0a47befef3b9ce737718d9587a59fc85.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0a47befef3b9ce737718d9587a59fc85.png\"><\/a><\/section>\n<section>MoE \u5c42\u4ec5\u9009\u62e9\u4e0e s_t \u2208 R^N_E \u4e2d top-K \u5143\u7d20\u76f8\u5bf9\u5e94\u7684 K \u4e2a\u4e13\u5bb6\uff08\u4ece N_E \u4e2d\uff09\uff0c\u6765\u4ea7\u751f\u5c42\u8f93\u51fa y_t \u2208 R^d_model\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0e315447fce37ef75400198f8e4bde36.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0e315447fce37ef75400198f8e4bde36.png\"><\/a><\/section>\n<section>\u521d\u6b65\u5b9e\u9a8c\u8868\u660e\uff0c\u03c3-MoE \u7684\u539f\u59cb<mark data-type=\"concepts\" data-id=\"c51052b5-4cd8-4df0-99bb-5aa643c2f027\">\u6b63\u5219\u5316<\/mark>\u5f80\u5f80\u4e0d\u7a33\u5b9a\uff0c\u6709\u65f6\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u671f\u95f4\u635f\u5931\u6fc0\u589e\u3002\u4e3a\u4e86\u907f\u514d\u8fd9\u79cd\u60c5\u51b5\uff0c\u7814\u7a76\u8005\u4ec5\u5728\u5e8f\u5217\u5185\u5e94\u7528<mark data-type=\"concepts\" data-id=\"c51052b5-4cd8-4df0-99bb-5aa643c2f027\">\u6b63\u5219\u5316<\/mark>\uff08\u800c\u4e0d\u662f<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u4e2d\u7684\u6240\u6709 token\uff09\u3002\u5bf9\u4e8e\u8f93\u5165\u5e8f\u5217 x_t\uff0ct \u2208 {1,&#8230;,T}\uff0c\u8ba1\u7b97\u5e73\u8861\u635f\u5931 L \u5982\u4e0b\u6240\u793a\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-fa2981fc6f77ea82826cb7ab9a62b7dc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-fa2981fc6f77ea82826cb7ab9a62b7dc.png\"><\/a><\/section>\n<section><strong>MoE <mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\u5c42<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5c06 MoE \u5f15\u5165<mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\u5c42\uff0c\u7814\u7a76\u8005\u5e94\u7528\u4e86 SwitchHead\uff0c\u5b83\u662f\u4e00\u79cd\u5c06 \u03c3-MoE \u6269\u5c55\u5230\u6ce8\u610f\u529b\u5c42\u7684 MoE \u65b9\u6cd5\u3002\u4e0e\u6807\u51c6\u591a\u5934\u6ce8\u610f\u529b\u5c42\u4e00\u6837\uff0cSwitchHead \u5c42\u4e2d\u7684\u6bcf\u4e2a\u5934\u5305\u542b\u56db\u4e2a\u8f6c\u6362\uff1a<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u3001\u952e\u3001\u503c\u548c\u8f93\u51fa\u6295\u5f71\u3002\u4f46\u662f\uff0cSwitchHead \u4f7f\u7528 MoE \u6765<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5316\u503c\u548c\u8f93\u51fa\u6295\u5f71\u3002<\/section>\n<section><\/section>\n<section>\u4e5f\u5373\uff0c\u6bcf\u4e2a\u5934\u90fd\u6709\u4e00\u4e2a\u4e0e\u4e4b\u5173\u8054\u7684<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u548c\u952e\u6295\u5f71\u4ee5\u53ca N_A \u503c\u548c\u8f93\u51fa\u6295\u5f71\uff0c\u5b83\u4eec\u9488\u5bf9\u6bcf\u4e2a\u8f93\u5165\u8fdb\u884c\u52a8\u6001\u9009\u62e9\u3002<\/section>\n<section><\/section>\n<section>\u952e\u548c<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u300c\u7167\u5e38\u300d\u8ba1\u7b97\uff1a\u7ed9\u5b9a\u4f4d\u7f6e t \u5904\u7684\u4e00\u4e2a\u8f93\u5165\uff0cx_t \u2208 R^d_model\uff0c\u5e76\u4e14 k^h_t = x_tW^h_K\u3001q^h_t = x_tW^h_Q\uff0ch \u2208{1,&#8230;,H} \u662f\u5934\u7d22\u5f15\u3002\u4e13\u5bb6\u5bf9\u8fd9\u4e9b\u503c\u7684\u9009\u62e9\u8ba1\u7b97\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d743a15c8ccd485ea7db91cf0ec62fe0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d743a15c8ccd485ea7db91cf0ec62fe0.png\"><\/a><\/section>\n<section>\u503c\u548c\u8f93\u51fa\u7684\u9009\u62e9\u662f\u72ec\u7acb\u7684\u3002\u8f93\u51fa\u7684\u9009\u62e9\u4f7f\u7528\u4e0d\u540c\u7684<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u77e9\u9635 W^h_SO \u2208 R^d_model\u00d7 N_A \u6765\u7c7b\u4f3c\u5730\u8ba1\u7b97\uff0c\u5373 <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0f9ecd02bb67ef6afd55c7b00ca04f28.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0f9ecd02bb67ef6afd55c7b00ca04f28.png\"><\/a> \u548c <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-411f257c5f65df282c45dbcdf2d80502.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-411f257c5f65df282c45dbcdf2d80502.png\"><\/a>\u3002\u8f93\u51fa y \u2208 R^d_model \u8ba1\u7b97\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d40433f30540aa889946c71b847e460c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d40433f30540aa889946c71b847e460c.png\"><\/a><\/section>\n<section><strong>\u5c42\u5206\u7ec4\uff1aMoE \u9ad8\u6548\u5c42\u5171\u4eab\u548c\u64cd\u4f5c\u5185\u7684\u5b50\u64cd\u4f5c<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u89c2\u5bdf\u5230\uff0c\u57fa\u4e8e MoE \u7684\u539f\u59cb UT \u53ea\u6709\u4e00\u4e2a\u5171\u4eab\u5c42\uff0c\u5728\u66f4\u5927\u89c4\u6a21\u4e0a\u5f80\u5f80\u96be\u4ee5\u83b7\u5f97\u826f\u597d\u7684\u6027\u80fd\u3002\u5047\u8bbe\u539f\u56e0\u6709\u4e8c\uff1a\u9996\u5148\uff0c\u968f\u7740\u7f51\u7edc\u89c4\u6a21\u7684\u6269\u5927\uff0c\u5c42\u4e2d\u4e13\u5bb6\u7684\u6570\u91cf\u4f1a\u8fc5\u901f\u589e\u52a0\uff0c\u4f46\u6211\u4eec\u65e0\u6cd5\u4ee5\u76f8\u540c\u7684\u901f\u5ea6\u589e\u52a0\u6d3b\u8dc3\u4e13\u5bb6 K \u6570\u91cf\u800c\u4e0d\u5927\u5e45\u589e\u52a0\u6240\u9700\u8ba1\u7b97\u91cf\u3002\u8fd9\u5c31\u8feb\u4f7f\u6211\u4eec\u964d\u4f4e\u6d3b\u8dc3\u4e13\u5bb6\u7684\u6bd4\u4f8b\uff0c\u800c\u8fd9\u901a\u5e38\u662f\u4e0d\u5229\u7684\u3002\u5176\u6b21\uff0c\u6ce8\u610f\u529b\u5934\u7684\u603b\u6570\u4fdd\u6301\u5728\u76f8\u5bf9\u8f83\u4f4e\u7684\u6c34\u5e73\uff0c\u8fd9\u5bf9\u4e8e\u4e00\u4e2a\u5927\u578b\u6a21\u578b\u6765\u8bf4\u53ef\u80fd\u662f\u4e0d\u591f\u7684\u3002\u589e\u52a0\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf\u4e5f\u540c\u6837\u4ee3\u4ef7\u9ad8\u6602\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u5728\u589e\u52a0\u6ce8\u610f\u529b\u5934\u603b\u6570\u7684\u540c\u65f6\uff0c\u53ef\u4ee5\u8c03\u7528\u4e00\u7ec4\u5c42\uff0c\u51cf\u5c11\u6bcf\u4e2a \u03c3-MoE \u4e2d\u7684\u4e13\u5bb6\u6570\u91cf\u3002\u6700\u7ec8\u7684\u7f51\u7edc\u662f\u901a\u8fc7\u91cd\u590d<mark data-type=\"tech_methods\" data-id=\"efb128f3-0b7c-45a8-a6ae-065fa93c6556\">\u5806\u53e0<\/mark>\u8fd9\u4e9b\u5171\u4eab\u76f8\u540c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684\u5c0f\u7ec4\u800c\u5f97\u5230\u7684\uff08\u4ece\u67d0\u79cd\u610f\u4e49\u4e0a\u8bf4\uff0c\u5c06\u7ec4\u91cd\u65b0\u5b9a\u4e49\u4e3a UT \u4e2d\u7684\u5171\u4eab\u5c42\uff09\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u56fe 1 \u63d0\u4f9b\u4e86\u4e00\u4e2a\u793a\u4f8b\uff0c\u6807\u8bb0\u4e3a\u300c\u5c42 A\u300d\uff08\u6216\u5c42 B\uff09\u7684\u6240\u6709\u5c42\u5728\u6574\u4e2a\u7f51\u7edc\u4e2d\u5171\u4eab\u76f8\u540c\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3002\u7ec4 G \u7684\u5927\u5c0f\u662f\u975e\u5171\u4eab\u5c42\u7684\u6570\u91cf\u3002\u5728\u7814\u7a76\u8005\u7684\u5b9e\u9a8c\u4e2d\uff0c\u7ec4\u5927\u5c0f\u5728 2 \u5230 4 \u4e4b\u95f4\uff0c\u5178\u578b\u7684\u5faa\u73af\u6b65\u9aa4\u6570\u4e3a 8 \u6216 9\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-30b337cee248cda0d445eba8cb44ec0b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-30b337cee248cda0d445eba8cb44ec0b.png\"><\/a><\/section>\n<section><strong>\u5728 UT \u4e2d\u6539\u8fdb\u4fe1\u53f7\u4f20\u64ad\u7684\u65b0\u578b<mark data-type=\"tech_methods\" data-id=\"edceaca3-7a3d-42b3-9189-d80b8b99a6ad\">\u5c42\u5f52\u4e00\u5316<\/mark>\u65b9\u6848<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u63d0\u51fa\u53e6\u4e00\u79cd\u65b9\u6cd5\u6765\u907f\u514d\u4e0a\u8ff0\u95ee\u9898\uff1a\u5728\u300c\u4e3b\u6570\u636e\u8def\u5f84\u300d\u4e2d\u4e0d\u4f7f\u7528\u5206<mark data-type=\"tech_methods\" data-id=\"edceaca3-7a3d-42b3-9189-d80b8b99a6ad\">\u5c42\u5f52\u4e00\u5316<\/mark>\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5bf9\u4e8e\u672c\u6587\u7684 UT\uff0c\u5728\u6ce8\u610f\u529b\u503c\u6295\u5f71\u4e4b\u524d\u4e0d\u4f7f\u7528\u5206\u5c42\u77e9\u9635\uff0c\u5728 \u03c3-MoE \u5c42\u4e4b\u524d\u4e5f\u4e0d\u4f7f\u7528\u5206\u5c42\u77e9\u9635\u3002\u76f8\u53cd\uff0c\u53ea\u6709\u5728\u7d27\u8ddf\u7740 sigmoid \u6216 softmax <mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u7684\u7ebf\u6027\u5c42\uff08\u5728\u8fd9\u4e9b\u975e\u7ebf\u6027\u5c42\u4e4b\u524d\u4ea7\u751f\u5173\u952e\u7684\u91cd\u5f52\u4e00\u5316\u6fc0\u6d3b\uff09\u4e4b\u524d\uff0c\u5373\u6ce8\u610f\u529b\u4e2d\u7684<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u548c\u5173\u952e\u6295\u5f71\u3001\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u5c42\u4e0a\u7684\u4e13\u5bb6\u9009\u62e9\u4ee5\u53ca\u6700\u7ec8\u5206\u7c7b\u5c42\u4e4b\u524d\uff0c\u624d\u4f1a\u4f7f\u7528 layernorm\u3002\u5982\u56fe 3 \u6240\u793a\u3002<\/section>\n<section><\/section>\n<section>\u7531\u4e8e\u524d\u9988\u5c42\u5185\u7684\u4e3b\u6570\u636e\u8def\u5f84\u4e0a\u53ea\u4f7f\u7528\u4e86 ReLU <mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\uff0c\u56e0\u6b64\u8f93\u51fa\u66f4\u65b0\u5c06\u4e0e\u8f93\u5165\u6210\u6b63\u6bd4\uff0c\u4ece\u800c\u6709\u6548\u5730\u89e3\u51b3\u4e86\u6b8b\u5dee\u589e\u957f\u95ee\u9898\uff0c\u540c\u65f6\u4e5f\u63d0\u4f9b\u4e86\u9ad8\u6548\u7684\u68af\u5ea6\u6d41\u8def\u5f84\u3002\u8fd9\u79cd\u65b9\u6848\u79f0\u4e3a \u300cperi-layernorm\u300d\u65b9\u6848\uff0c\u5b83\u4ecb\u4e8e\u300cpre-layernorm\u300d\u548c\u300cpost-layernorm\u300d\u65b9\u6848\u4e4b\u95f4\uff0c\u5c06 layernorm \u5b9a\u4f4d\u5728\u6b8b\u5dee\u8fde\u63a5\u7684\u300c\u5468\u56f4\u300d\uff08\u4f46\u4e0d\u5728\u5176\u4e0a\uff09\u3002\u5177\u4f53\u5982\u4e0b\u56fe 3 \u6240\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8bf6db3d349fdac15962640bd0b65df2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8bf6db3d349fdac15962640bd0b65df2.png\"><\/a><\/section>\n<section><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u8bba\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u5c55\u793a\u4e86\u4f7f\u7528\u6d41\u884c\u7684 C4 \u6570\u636e\u96c6\u8fdb\u884c\u8bed\u8a00\u5efa\u6a21\u65f6 MoEUT \u6027\u80fd\u548c\u6548\u7387\u7684\u4e3b\u8981\u5b9e\u9a8c\u7ed3\u679c\u3002\u4e3a\u4e86\u8bc1\u660e MoEUT \u7684\u901a\u7528\u6027\uff0c\u7814\u7a76\u8005\u8fd8\u5c55\u793a\u4e86\u5728 SlimPajama \u548c peS2o \u8bed\u8a00\u5efa\u6a21\u6570\u636e\u96c6\u4e0a\u7684\u4e3b\u8981\u7ed3\u679c\uff0c\u4ee5\u53ca\u5728 \u300cThe Stack\u300d\u4e0a\u7684\u4ee3\u7801\u751f\u6210\u3002<\/section>\n<section><\/section>\n<section><strong>\u4e0e\u6807\u51c6 Transformer \u5bf9\u6bd4 Scaling<\/strong><\/section>\n<section><\/section>\n<section>MoEUT \u7684 Scaling \u7ed3\u679c\u5982\u56fe 4 \u6240\u793a\u3002y \u8f74\u663e\u793a\u7684\u662f C4 held-out \u5b50\u96c6\u4e0a\u7684\u590d\u6742\u5ea6\u3002\u5728<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6570\u91cf\u76f8\u540c\u7684\u60c5\u51b5\u4e0b\uff0cMoEUT \u6a21\u578b\u7565\u5fae\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b\uff08\u56fe 4a\uff09\uff0c\u800c\u4e14\u968f\u7740\u89c4\u6a21\u7684\u6269\u5927\uff0c\u5dee\u8ddd\u5448\u6269\u5927\u8d8b\u52bf\u3002\u7814\u7a76\u8005\u8fd8\u4e0e\u975e\u5171\u4eab \u03c3-MoE \u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u8be5\u6a21\u578b\u7684\u8868\u73b0\u660e\u663e\u4e0d\u5982 MoEUT\uff0c\u8fd9\u8868\u660e\u5171\u4eab\u5c42\u5177\u6709\u660e\u663e\u7684\u4f18\u52bf\u3002\u6b64\u5916\u5982\u56fe 4b \u663e\u793a\uff0c\u5c31\u8bad\u7ec3\u671f\u95f4\u6240\u6709\u524d\u5411\u4f20\u9012\u6240\u82b1\u8d39\u7684\u603b MAC \u8fd0\u7b97\u6b21\u6570\u800c\u8a00\uff0cMoEUT \u8fdc\u8fdc\u4f18\u4e8e\u57fa\u7ebf\u5bc6\u96c6\u6a21\u578b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-bbf4e03df5616b2a32ff9a41bce000e9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-bbf4e03df5616b2a32ff9a41bce000e9.png\"><\/a><\/section>\n<section><strong>\u4ee3\u7801\u751f\u6210\u6027\u80fd<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u8bc1\u5b9e MoEUT \u5728\u4e0d\u540c\u4efb\u52a1\u9886\u57df\u7684\u6709\u6548\u6027\uff0c\u7814\u7a76\u8005\u5728\u300cThe Stack\u300d\u6570\u636e\u96c6\u7684\u4e00\u4e2a\u5b50\u96c6\u4e0a\u5bf9\u5176\u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u8be5\u6570\u636e\u96c6\u662f\u4e00\u4e2a\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u3002\u7531\u4e8e\u65e0\u6cd5\u8fdb\u884c\u5b8c\u6574\u7684 epoch \u8bad\u7ec3\uff0c\u56e0\u6b64\u8fd9\u91cc\u53ea\u4f7f\u7528\u4e86\u51e0\u79cd\u8bed\u8a00\u5e76\u6df7\u5408\u4f7f\u7528\u4e86\u8fd9\u4e9b\u8bed\u8a00\uff1aPython\u3001HTML\u3001C++\u3001Rust\u3001JavaScript\u3001Haskell\u3001Scala \u548c\u6c47\u7f16\u3002\u7814\u7a76\u8005\u5728\u6570\u636e\u96c6\u7684\u4e00\u4e2a held-out \u5b50\u96c6\u4e0a\u8bc4\u4f30\u4e86 MoEUT\u3002\u7ed3\u679c\u5982\u56fe 5 \u6240\u793a\uff0c\u4e0e\u81ea\u7136\u8bed\u8a00\u9886\u57df\u7684\u7ed3\u679c\u4e00\u81f4\uff1aMoEUT \u7684\u8868\u73b0\u4f18\u4e8e\u57fa\u7ebf\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95a0bc7f25b71aabfc3d4b6dd1be6f02.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95a0bc7f25b71aabfc3d4b6dd1be6f02.png\"><\/a><\/section>\n<section>\u56fe 6 \u5c55\u793a\u4e86 layer grouping \u5bf9 244M <mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark> MoEUT \u6a21\u578b\u7684\u5f71\u54cd\u3002\u7814\u7a76\u8005\u53d1\u73b0 G = 2 \u662f\u6700\u4f73\u503c\uff0c\u800c\u4e14\u5c42\u7ef4\u5ea6\u7684\u5faa\u73af\u786e\u5b9e\u662f\u6709\u76ca\u7684\u3002<\/section>\n<section><\/section>\n<section><strong>\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u96f6\u6837\u672c\u8868\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u8bc4\u4f30\u4e86 MoEUT \u5728\u516d\u4e2a\u4e0d\u540c\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7684\u96f6\u6837\u672c\u6027\u80fd\uff1aLAMBADA\u3001BLiMP\u3001Children&#8217;s Book Test (CBT) \u3001HellaSwag\u3001PIQA \u548c ARC-E\u3002\u7ed3\u679c\u89c1\u8868 1\uff0cMoEUT \u7684\u8868\u73b0\u5f80\u5f80\u4f18\u4e8e\u57fa\u7ebf\uff0c\u4f46\u5728\u6240\u6709\u60c5\u51b5\u4e0b\u5dee\u5f02\u90fd\u5f88\u5c0f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-f36f4df6e3773433b330c0b2ae077357.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-f36f4df6e3773433b330c0b2ae077357.png\"><\/a><\/section>\n<section>\u7814\u7a76\u8005\u8fd8\u5c06 MoEUT \u4e0e\u53e6\u4e00\u4e2a<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6a21\u578b Sparse Universal Transformer\uff08SUT\uff09\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0cSUT \u662f\u6700\u8fd1\u63d0\u51fa\u7684\u4e00\u4e2a UT \u6a21\u578b\uff0c\u4e5f\u4f7f\u7528\u4e86 MoE \u5c42\uff0c\u4e14\u4ee5\u524d\u672a\u5728\u6807\u51c6\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e2d\u8fdb\u884c\u8fc7\u8bc4\u4f30\u3002\u867d\u7136 MoEUT \u548c SUT \u90fd\u5728\u524d\u9988\u5c42\u548c\u6ce8\u610f\u529b\u5c42\u4f7f\u7528\u4e86 MoE\uff0c\u4f46\u8fd9\u4e24\u79cd\u65b9\u6cd5\u5728\u4e0d\u540c\u5c42\u9762\u4e0a\u5b58\u5728\u4e00\u4e9b\u6280\u672f\u5dee\u5f02\uff1aSUT \u4f7f\u7528\u7ade\u4e89\u6027\u4e13\u5bb6\u9009\u62e9\uff08softmax\uff09\u3001\u591a\u91cd\u8d1f\u8f7d\u5e73\u8861\u635f\u5931\u548c\u66f4\u5927\u7684\u4e13\u5bb6\u89c4\u6a21\uff0c\u4e14\u91c7\u7528 post-layernorm \u6a21\u5f0f\uff0c\u4e0d\u4f7f\u7528 layer grouping\u3002\u4e0e MoEUT \u7684\u65b9\u6cd5\u4e0d\u540c\uff0cSUT \u5728\u5c42\u7ef4\u5ea6\u4e0a\u4f7f\u7528\u4e86\u81ea\u9002\u5e94\u8ba1\u7b97\u65f6\u95f4\uff08ACT\uff09\u3002<\/section>\n<section><\/section>\n<section>\u7ed3\u679c\u5982\u56fe 7 \u6240\u793a\u3002\u4e0e MoEUT \u548c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5339\u914d\u7684\u5bc6\u96c6\u57fa\u7ebf\u76f8\u6bd4\uff0c SUT \u5728\u6027\u80fd\u4e0a\u6709\u660e\u663e\u7684\u52a3\u52bf\u3002\u7814\u7a76\u8005\u8ba4\u4e3a\u8fd9\u79cd\u6027\u80fd\u4f4e\u4e0b\u7684\u4e3b\u8981\u539f\u56e0\u662f\u4f5c\u8005\u5c06 ACT \u673a\u5236\u4f5c\u4e3a\u5176\u6a21\u578b\u7684\u4e3b\u8981\u7ec4\u6210\u90e8\u5206\u4e4b\u4e00\u3002\u79fb\u9664 ACT \u540e\uff0c\u6027\u80fd\u663e\u8457\u63d0\u9ad8\u3002\u7136\u800c\uff0c\u5373\u4f7f\u5728\u8fd9\u79cd\u8bbe\u7f6e\u4e0b\uff0c\u5b83\u7684\u6027\u80fd\u4ecd\u7136\u4f4e\u4e8e MoEUT \u548c\u6807\u51c6 Transformer \u57fa\u7ebf\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-207067c0c0ba1ecfe1a00993fb7bdd8e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-207067c0c0ba1ecfe1a00993fb7bdd8e.png\"><\/a><\/section>\n<section>\u7814\u7a76\u8005\u8fd8\u5bf9\u300cperi &#8211; <mark data-type=\"tech_methods\" data-id=\"edceaca3-7a3d-42b3-9189-d80b8b99a6ad\">\u5c42\u5f52\u4e00\u5316<\/mark>\u300d\u8fdb\u884c\u4e86\u8bc4\u4f30\u3002\u56fe 8 \u663e\u793a\u4e86\u7ed3\u679c\u3002\u672c\u6587\u7684<mark data-type=\"tech_methods\" data-id=\"edceaca3-7a3d-42b3-9189-d80b8b99a6ad\">\u5c42\u5f52\u4e00\u5316<\/mark>\u65b9\u6848\u59cb\u7ec8\u8868\u73b0\u6700\u4f73\u3002\u5c0f\u6a21\u578b\u7684\u5dee\u8ddd\u66f4\u5927\uff0c\u800c\u5927\u6a21\u578b\u7684\u5dee\u8ddd\u5219\u8d8a\u6765\u8d8a\u5c0f\uff08\u5bf9\u4e8e 719M <mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6a21\u578b\uff0cperi-norm \u548c post-norm \u4e4b\u95f4\u7684\u5dee\u8ddd\u5fae\u4e4e\u5176\u5fae\uff09\u3002\u540c\u65f6\uff0c\u968f\u7740\u8bad\u7ec3\u6b65\u6570\u7684\u589e\u52a0\uff0cperi-norm \u548c post-norm \u4e4b\u95f4\u7684\u5dee\u8ddd\u4e5f\u5728\u589e\u5927\uff0c\u56e0\u6b64\u5982\u679c\u6a21\u578b\u7684\u8bad\u7ec3\u65f6\u95f4\u66f4\u957f\uff0c\u5c31\u6709\u53ef\u80fd\u83b7\u5f97\u66f4\u9ad8\u7684\u6536\u76ca\u3002<\/section>\n<section><\/section>\n<section><strong>\u8c03\u6574\u4e13\u5bb6\u9009\u62e9\u673a\u5236<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u66f4\u597d\u5730\u7406\u89e3 MoEUT \u7684\u4e13\u5bb6\u9009\u62e9\uff0c\u7814\u7a76\u8005\u5206\u6790\u4e86\u5728 C4 \u4e0a\u8bad\u7ec3\u7684 244M <mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark> MoEUT \u6a21\u578b\u7684 MLP \u5757\u4e2d\u7684\u4e13\u5bb6\u9009\u62e9\u3002\u672c\u8282\u4e2d\u7684\u6240\u6709\u5b9e\u9a8c\u90fd\u662f\u901a\u8fc7\u8ba1\u7b97 C4 <mark data-type=\"concepts\" data-id=\"698eae06-517e-4332-ba61-2de17c83ffb6\">\u9a8c\u8bc1\u96c6<\/mark>\u4e0a G = 2\uff08\u5373\u6a21\u578b\u7ec4\u4e2d\u6709\u4e24\u5c42\uff09\u6a21\u578b\u7684\u7edf\u8ba1\u6570\u636e\u8fdb\u884c\u7684\u3002\u8fd9\u91cc\u53ea\u5c55\u793a\u4e86\u6a21\u578b\u7ec4\u7b2c\u4e00\u5c42\u7684\u884c\u4e3a\uff0c\u56e0\u4e3a\u7814\u7a76\u8005\u53d1\u73b0\u7b2c\u4e8c\u5c42\u7684\u7ed3\u679c\u5728\u672c\u8d28\u4e0a\u662f\u76f8\u4f3c\u7684\u3002<\/section>\n<section><\/section>\n<section>\u7ed3\u679c\u8868\u660e\uff0cMoEUT \u80fd\u591f\u6839\u636e\u4e0d\u540c\u60c5\u51b5\u52a8\u6001\u8c03\u6574\u5176\u4e13\u5bb6\u9009\u62e9\u673a\u5236\u3002\u6709\u65f6\uff0c\u4e13\u5bb6\u4f1a\u88ab\u5206\u914d\u7ed9\u6d41\u884c\u7684 token\uff0c\u800c\u5728\u5176\u4ed6\u60c5\u51b5\u4e0b\uff0c\u4e13\u5bb6\u4f1a\u5728\u5404\u5c42\u4e4b\u95f4\u5171\u4eab\u6216\u4e13\u95e8\u5316\uff0c\u8fd9\u53d6\u51b3\u4e8e\u54ea\u79cd\u65b9\u5f0f\u66f4\u9002\u5408\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u4e13\u5bb6\u7684\u8de8\u5c42\u4f7f\u7528\u3002\u5982\u56fe 9 \u6240\u793a\uff0c\u53f3\u4e0b\u89d2\u7684\u9ec4\u70b9\u8868\u793a\u4e00\u4e9b\u4e13\u5bb6\u4e3b\u8981\u88ab\u5206\u914d\u5230\u6700\u540e\u4e00\u5c42\u3002\u7136\u800c\uff0c\u5bf9\u4e8e\u5176\u4ed6\u4e13\u5bb6\u6765\u8bf4\uff0c\u4e13\u5bb6\u88ab\u6fc0\u6d3b\u7684\u5c42\u8303\u56f4\u5f88\u5e7f\u3002\u4e13\u5bb6\u4f3c\u4e4e\u662f\u5728\u8fde\u7eed\u7684\u5c42\u5e8f\u5217\u4e2d\u88ab\u6fc0\u6d3b\u7684\uff0c\u8fd9\u53ef\u4ee5\u4ece\u7eb5\u5411\u6392\u5217\u7684\u5bbd\u9614\u7ed3\u6784\u4e2d\u770b\u51fa\u3002\u56e0\u6b64\u53ef\u4ee5\u5f97\u51fa\u8fd9\u6837\u7684\u7ed3\u8bba\uff1a\u5982\u679c\u6709\u5fc5\u8981\uff0cMoEUT \u80fd\u591f\u4e13\u6ce8\u4e8e\u7279\u5b9a\u5c42\uff0c\u5e76\u53ef\u5728\u5404\u5c42\u4e4b\u95f4\u5171\u4eab<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-1ac5c5d4ee469aec663ba1c4be256a28.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-1ac5c5d4ee469aec663ba1c4be256a28.png\"><\/a><\/section>\n<section>\u6bcf\u4e2a token \u4e13\u5bb6\u9009\u62e9\u7684\u591a\u6837\u6027\u3002\u7814\u7a76\u8005\u5206\u6790\u4e86 MLP \u5404\u5c42\u9488\u5bf9\u7ed9\u5b9a\u8f93\u5165 token \u5728\u4e0d\u540c\u5c42\u548c\u4e0a\u4e0b\u6587\u4e2d\u7684\u4e13\u5bb6\u9009\u62e9\u591a\u6837\u6027\u3002\u4e3a\u6b64\uff0c\u4ed6\u4eec\u6d4b\u91cf\u4e86\u4e0d\u540c\u5c42\u5728\u4e0d\u540c\u4f4d\u7f6e \/ \u4e0a\u4e0b\u6587\u4e0b\u4e3a\u5355\u4e2a token \u6fc0\u6d3b\u7684\u4e13\u5bb6\u603b\u6570\u3002\u7ed3\u679c\u5982\u56fe 10 \u6240\u793a\u3002<\/section>\n<section><\/section>\n<section>\u5404\u680f\/\u4f4d\u7f6e\u7684\u4e13\u5bb6\u9009\u62e9\u52a8\u6001\u3002\u7ed3\u679c\u5982\u56fe 11 \uff0c\u5728\u540e\u7eed\u5c42\u4e2d\uff0c\u6240\u9009\u4e13\u5bb6\u4e4b\u95f4\u5b58\u5728\u4e0d\u53ef\u5ffd\u7565\u7684\u91cd\u53e0\uff1b\u4f46\u662f\uff0c\u8fd9\u79cd\u91cd\u53e0\u8fd8\u8fdc\u8fdc\u6ca1\u6709\u8fbe\u5230\u5b8c\u5168\u91cd\u53e0\u7684\u7a0b\u5ea6\u3002\u8fd9\u8868\u660e\uff0c\u4e13\u5bb6\u901a\u5e38\u5728\u5355\u5217\u4e2d\u52a8\u6001\u53d8\u5316\uff0c\u5728\u4e0d\u540c\u5c42\u4e2d\u6267\u884c\u4e0d\u540c\u7684\u529f\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95160e75b6022e1d59f4f952ea75d38a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95160e75b6022e1d59f4f952ea75d38a.png\"><\/a><\/section>\n<section>\u66f4\u591a\u7814\u7a76\u7ec6\u8282\uff0c\u53ef\u53c2\u8003\u539f\u8bba\u6587\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-19-2\" target=\"_blank\">Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:Jurgen\u3001\u66fc\u5b81\u7b49\u5927 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","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":"default","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":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-36209","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=36209\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7\",\"datePublished\":\"2025-01-07T07:00:12+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209\"},\"wordCount\":304,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-82efdcb624831ee3a6b4b43824021658.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=36209#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=36209\",\"name\":\"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-82efdcb624831ee3a6b4b43824021658.png\",\"datePublished\":\"2025-01-07T07:00:12+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=36209\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-82efdcb624831ee3a6b4b43824021658.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-82efdcb624831ee3a6b4b43824021658.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36209#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=36209","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=36209#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=36209"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7","datePublished":"2025-01-07T07:00:12+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=36209"},"wordCount":304,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=36209#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=36209#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36209","url":"https:\/\/17aitech.com\/?p=36209","name":"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36209#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36209#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png","datePublished":"2025-01-07T07:00:12+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36209#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36209"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36209#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-82efdcb624831ee3a6b4b43824021658.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36209#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"Jurgen\u3001\u66fc\u5b81\u7b49\u5927\u4f6c\u65b0\u4f5c\uff1aMoE\u91cd\u58516\u5e74\u524d\u7684Universal Transformer\uff0c\u9ad8\u6548\u5347\u7ea7"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36209","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=36209"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36209\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36209"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36209"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36209"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}