{"id":34758,"date":"2024-12-12T11:57:43","date_gmt":"2024-12-12T03:57:43","guid":{"rendered":"https:\/\/17aitech.com\/?p=34758"},"modified":"2024-12-12T11:57:43","modified_gmt":"2024-12-12T03:57:43","slug":"%e7%ae%97%e6%b3%95%e3%80%81%e7%b3%bb%e7%bb%9f%e5%92%8c%e5%ba%94%e7%94%a8%ef%bc%8c%e4%b8%89%e4%b8%aa%e8%a7%86%e8%a7%92%e5%85%a8%e9%9d%a2%e8%af%bb%e6%87%82%e6%b7%b7%e5%90%88%e4%b8%93%e5%ae%b6%ef%bc%88mo","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34758","title":{"rendered":"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-26-5\" target=\"_blank\">\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"63\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<section>LLM \u5f88\u5f3a\uff0c\u800c\u4e3a\u4e86\u5b9e\u73b0 LLM \u7684\u53ef\u6301\u7eed\u6269\u5c55\uff0c\u6709\u5fc5\u8981\u627e\u5230\u5e76\u5b9e\u73b0\u80fd\u63d0\u5347\u5176\u6548\u7387\u7684\u65b9\u6cd5\uff0c\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u5c31\u662f\u8fd9\u7c7b\u65b9\u6cd5\u7684\u4e00\u5927\u91cd\u8981\u6210\u5458\u3002<\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n<p>\u6700\u8fd1\uff0c\u5404\u5bb6\u79d1\u6280\u516c\u53f8\u63d0\u51fa\u7684\u65b0\u4e00\u4ee3\u5927\u6a21\u578b\u4e0d\u7ea6\u800c\u540c\u5730\u6b63\u5728\u4f7f\u7528\u6df7\u5408\u4e13\u5bb6\uff08Mixture of Experts\uff1aMoE\uff09\u65b9\u6cd5\u3002<\/p>\n<p>\u6df7\u5408\u4e13\u5bb6\u8fd9\u4e00\u6982\u5ff5\u6700\u65e9\u8bde\u751f\u4e8e 1991 \u5e74\u7684\u8bba\u6587\u300aAdaptive mixtures of local experts\u300b\uff0c\u4e09\u5341\u591a\u5e74\u6765\u5df2\u7ecf\u5f97\u5230\u4e86\u5e7f\u6cdb\u7684\u63a2\u7d22\u548c\u53d1\u5c55\u3002\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u7a00\u758f\u95e8\u63a7 MoE \u7684\u51fa\u73b0\u548c\u53d1\u5c55\uff0c\u5c24\u5176\u662f\u4e0e\u57fa\u4e8e Transformer \u7684\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u76f8\u7ed3\u5408\uff0c\u8fd9\u79cd\u5df2\u6709\u4e09\u5341\u591a\u5e74\u5386\u53f2\u7684\u6280\u672f\u7115\u53d1\u51fa\u4e86\u65b0\u7684\u751f\u673a\u3002<\/p>\n<p>MoE \u6846\u67b6\u57fa\u4e8e\u4e00\u4e2a\u7b80\u5355\u5374\u53c8\u5f3a\u5927\u601d\u60f3\uff1a\u6a21\u578b\u7684\u4e0d\u540c\u90e8\u5206\uff08\u79f0\u4e3a\u4e13\u5bb6\uff09\u4e13\u6ce8\u4e8e\u4e0d\u540c\u7684\u4efb\u52a1\u6216\u6570\u636e\u7684\u4e0d\u540c\u65b9\u9762\u3002<\/p>\n<p>\u4f7f\u7528\u8fd9\u4e00\u8303\u5f0f\u65f6\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u8f93\u5165\uff0c\u4ec5\u6709\u4e0e\u4e4b\u76f8\u5173\u7684\u4e13\u5bb6\uff08Expert\uff09\u624d\u4f1a\u53c2\u4e0e\u5904\u7406\uff0c\u8fd9\u6837\u4e00\u6765\u4fbf\u80fd\u63a7\u5236\u8ba1\u7b97\u6210\u672c\uff0c\u540c\u65f6\u4ecd\u80fd\u53d7\u76ca\u4e8e\u5927\u91cf\u4e13\u4e1a\u77e5\u8bc6\u3002\u56e0\u6b64\uff0cMoE \u53ef\u5728\u4e0d\u5927\u5e45\u63d0\u5347\u8ba1\u7b97\u9700\u6c42\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u5927<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u80fd\u529b\u3002<\/p>\n<p>\u5982\u56fe 1 \u6240\u793a\uff0cMoE \u76f8\u5173\u7814\u7a76\u589e\u957f\u5f3a\u52b2\uff0c\u5c24\u5176\u662f\u5728 2024 \u5e74 Mixtral-8x7B \u4ee5\u53ca Grok-1\u3001DBRX\u3001Arctic\u3001DeepSeek-V2 \u7b49\u5404\u79cd\u4ea7\u4e1a\u7ea7 LLM \u51fa\u73b0\u4e4b\u540e\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png\"><\/a><\/p>\n<p>\u8fd9\u5f20\u56fe\u6765\u81ea\u9999\u6e2f\u79d1\u6280\u5927\u5b66\uff08\u5e7f\u5dde\uff09\u7684\u4e00\u4e2a\u7814\u7a76\u56e2\u961f\u8fd1\u65e5\u53d1\u5e03\u7684\u4e00\u7bc7 MoE \u7efc\u8ff0\u62a5\u544a\uff0c\u5176\u4e2d\u6e05\u6670\u4e14\u5168\u9762\u5730\u603b\u7ed3\u4e86 MoE \u76f8\u5173\u7814\u7a76\uff0c\u5e76\u63d0\u51fa\u4e86\u4e00\u79cd\u5168\u65b0\u7684\u5206\u7c7b\u6cd5\uff0c\u5c06\u8fd9\u4e9b\u7814\u7a76\u5f52\u7c7b\u5230\u4e86\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\u4e09\u5927\u7c7b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-97dd7d6abb9af5e9c5cfc7d91c1d0a53.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-97dd7d6abb9af5e9c5cfc7d91c1d0a53.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aA Survey on Mixture of Experts<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2407.06204<\/p>\n<\/li>\n<\/ul>\n<p><mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u6574\u7406\u4e86\u8fd9\u7bc7\u7efc\u8ff0\u62a5\u544a\u7684\u5185\u5bb9\u4e3b\u5e72\uff0c\u4ee5\u5e2e\u52a9\u8bfb\u8005\u4e86\u89e3\u5f53\u524d MoE \u7684\u53d1\u5c55\u6982\u51b5\uff0c\u66f4\u591a\u8be6\u60c5\u8bf7\u9605\u8bfb\u539f\u8bba\u6587\u3002\u6b64\u5916\uff0c\u6211\u4eec\u4e5f\u5728\u6587\u672b\u6574\u7406\u4e86\u4e00\u4e9b\u4e0e MoE \u76f8\u5173\u7684\u62a5\u9053\u3002<\/p>\n<p><strong>\u6df7\u5408\u4e13\u5bb6\u7684\u80cc\u666f\u77e5\u8bc6<\/strong><\/p>\n<p>\u5728\u57fa\u4e8e Transformer \u7684\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u4e2d\uff0c\u6bcf\u4e2a\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u5c42\u7684\u7ec4\u6210\u5f62\u5f0f\u901a\u5e38\u662f \ud835\udc41 \u4e2a\u300c<mark data-type=\"tech_methods\" data-id=\"9fbe8681-737e-4466-9126-9e37fc194ec6\">\u4e13\u5bb6\u7f51\u7edc<\/mark>\u300d{\ud835\udc53_1, &#8230; , \ud835\udc53_\ud835\udc41} \u642d\u914d\u4e00\u4e2a\u300c\u95e8\u63a7\u7f51\u7edc\u300dG\u3002<\/p>\n<p>\u8fd9\u4e2a\u95e8\u63a7\u7f51\u7edc\u7684\u5f62\u5f0f\u901a\u5e38\u662f\u4e00\u4e2a\u4f7f\u7528 softmax <mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u7684\u7ebf\u6027\u7f51\u7edc\uff0c\u5176\u4f5c\u7528\u662f\u5c06\u8f93\u5165\u5f15\u5bfc\u81f3\u5408\u9002\u7684<mark data-type=\"tech_methods\" data-id=\"9fbe8681-737e-4466-9126-9e37fc194ec6\">\u4e13\u5bb6\u7f51\u7edc<\/mark>\u3002MoE \u5c42\u7684\u653e\u7f6e\u4f4d\u7f6e\u662f\u5728 Transformer \u6a21\u5757\u5185\uff0c\u4f5c\u7528\u662f\u9009\u53d6\u524d\u5411\u7f51\u7edc\uff08FFN\uff09\uff0c\u901a\u5e38\u4f4d\u4e8e<mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\uff08SA\uff09\u5b50\u5c42\u4e4b\u540e\u3002\u8fd9\u79cd\u653e\u7f6e\u65b9\u5f0f\u5f88\u5173\u952e\uff0c\u56e0\u4e3a\u968f\u7740\u6a21\u578b\u589e\u5927\uff0cFFN \u7684\u8ba1\u7b97\u9700\u6c42\u4e5f\u4f1a\u589e\u52a0\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\u8fbe\u5230 5400 \u4ebf\u7684 PaLM \u6a21\u578b\u4e2d\uff0c90% \u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u90fd\u4f4d\u4e8e\u5176 FFN \u5c42\u4e2d\u3002<\/p>\n<p>\u7528\u6570\u5b66\u5f62\u5f0f\u63cf\u8ff0\u7684\u8bdd\uff1a\u6bcf\u4e2a<mark data-type=\"tech_methods\" data-id=\"9fbe8681-737e-4466-9126-9e37fc194ec6\">\u4e13\u5bb6\u7f51\u7edc<\/mark> \ud835\udc53_\ud835\udc56 \uff08\u901a\u5e38\u662f\u4e00\u4e2a\u7ebf\u6027 &#8211; ReLU &#8211; \u7ebf\u6027\u7f51\u7edc\uff09\u90fd\u7531 W_\ud835\udc56 \u8fdb\u884c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5316\uff0c\u5176\u63a5\u6536\u540c\u4e00\u8f93\u5165 x \u5e76\u751f\u6210\u8f93\u51fa \ud835\udc53_\ud835\udc56 (x; W_\ud835\udc56)\u3002\u540c\u65f6\uff0c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u4e3a \u0398 \u7684\u95e8\u63a7\u7f51\u7edc G\uff08\u901a\u5e38\u7531\u4e00\u4e2a\u7ebf\u6027 &#8211; ReLU &#8211; \u7ebf\u6027 &#8211; softmax \u7f51\u7edc\u6784\u6210\uff09\u5f97\u5230\u8f93\u51fa G (x; \u0398)\u3002\u518d\u6839\u636e\u95e8\u63a7\u51fd\u6570\u7684\u8bbe\u8ba1\u65b9\u5f0f\uff0c\u53ef\u4ee5\u5c06 MoE \u5c42\u5927\u81f4\u5206\u4e3a\u4ee5\u4e0b\u4e24\u7c7b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8cb0c23743e17f294773da2fe7131372.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8cb0c23743e17f294773da2fe7131372.png\"><\/a><\/p>\n<p><strong>\u5bc6\u96c6 MoE<\/strong><\/p>\n<p>\u5bc6\u96c6\u6df7\u5408\u4e13\u5bb6\u5c42\u662f\u5728\u6bcf\u6b21\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\u6fc0\u6d3b\u6240\u6709<mark data-type=\"tech_methods\" data-id=\"9fbe8681-737e-4466-9126-9e37fc194ec6\">\u4e13\u5bb6\u7f51\u7edc<\/mark> {\ud835\udc53_1, &#8230; , \ud835\udc53_\ud835\udc41}\u3002\u65e9\u671f\u7684 MoE \u7814\u7a76\u666e\u904d\u91c7\u7528\u4e86\u8fd9\u4e00\u7b56\u7565\u3002\u8fd1\u6bb5\u65f6\u95f4\u4e5f\u6709\u4e00\u4e9b\u7814\u7a76\u91c7\u7528\u4e86\u5bc6\u96c6 MoE\uff0c\u6bd4\u5982 EvoMoE\u3001MoLE \u3001LoRAMoE \u548c DS-MoE\u3002\u56fe 2a \u7ed9\u51fa\u4e86\u5bc6\u96c6 MoE \u5c42\u7684\u7ed3\u6784\u3002\u56e0\u6b64\uff0c\u5bc6\u96c6 MoE \u5c42\u7684\u8f93\u51fa\u53ef\u4ee5\u8868\u793a\u6210\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7babdfbf0a87588f085cbc4b4cedbac2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7babdfbf0a87588f085cbc4b4cedbac2.png\"><\/a><\/p>\n<p>\u5176\u4e2d\uff0c\ud835\udc54(x; \u0398) \u662f softmax \u8fd0\u7b97\u4e4b\u524d\u7684\u95e8\u63a7\u503c\u3002<\/p>\n<p><strong>\u7a00\u758f MoE<\/strong><\/p>\n<p>\u5c3d\u7ba1\u5bc6\u96c6\u6df7\u5408\u4e13\u5bb6\u7684\u9884\u6d4b\u51c6\u786e\u5ea6\u901a\u5e38\u66f4\u9ad8\uff0c\u4f46\u5176\u8ba1\u7b97\u8d1f\u8f7d\u4e5f\u975e\u5e38\u9ad8\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cShazeer et al. \u7684\u8bba\u6587\u300aOutrageously large neural networks: The sparsely-gated mixture-of-experts layer\u300b\u5f15\u5165\u4e86\u7a00\u758f\u95e8\u63a7 MoE \u5c42\uff0c\u5176\u80fd\u5728\u6bcf\u6b21\u524d\u5411\u901a\u8fc7\u65f6\u4ec5\u6fc0\u6d3b\u9009\u5b9a\u7684\u4e13\u5bb6\u5b50\u96c6\u3002\u8be5\u7b56\u7565\u5b9e\u73b0\u7a00\u758f\u6027\u7684\u65b9\u5f0f\u662f\u8ba1\u7b97 top-k \u4e2a\u4e13\u5bb6\u7684\u8f93\u51fa\u7684\u52a0\u6743\u548c\uff0c\u800c\u975e\u5c06\u6240\u6709\u4e13\u5bb6\u7684\u8f93\u51fa\u805a\u5408\u5230\u4e00\u8d77\u3002\u56fe 2b \u5c55\u793a\u4e86\u8fd9\u79cd\u7a00\u758f MoE \u5c42\u7684\u7ed3\u6784\u3002<\/p>\n<p>\u6839\u636e\u4e0a\u8ff0\u8bba\u6587\u63d0\u51fa\u7684\u6846\u67b6\uff0c\u53ef\u5bf9 2.2 \u5f0f\u8fdb\u884c\u4fee\u6539\u4ee5\u53cd\u6620\u7a00\u758f\u95e8\u63a7\u673a\u5236\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-40a1723bfb8fb84b5fed8e9d5dce5dc2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-40a1723bfb8fb84b5fed8e9d5dce5dc2.png\"><\/a><\/p>\n<p>\u8fd9\u91cc\u89e3\u91ca\u4e00\u4e0b\uff1aTopK (\u30fb, \ud835\udc58) \u51fd\u6570\u662f\u4ec5\u4fdd\u7559\u5411\u91cf\u539f\u59cb\u503c\u7684\u524d k \u9879\uff0c\u540c\u65f6\u5c06\u5176\u5b83\u9879\u8bbe\u7f6e\u4e3a \u2212\u221e\u3002\u4e4b\u540e\u662f softmax \u8fd0\u7b97\uff0c\u6240\u6709 \u2212\u221e \u9879\u90fd\u4f1a\u53d8\u6210\u8fd1\u4f3c\u4e8e\u96f6\u3002\u8d85<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark> k \u8981\u6839\u636e\u5177\u4f53\u5e94\u7528\u9009\u53d6\uff0c\u5e38\u89c1\u9009\u9879\u662f \ud835\udc58 = 1 \u6216 \ud835\udc58 = 2\u3002\u52a0\u5165\u566a\u58f0\u9879 R_noise \u662f\u8bad\u7ec3\u7a00\u758f\u95e8\u63a7 MoE \u5c42\u7684\u4e00\u79cd\u5e38\u7528\u7b56\u7565\uff0c\u53ef\u4fc3\u8fdb\u4e13\u5bb6\u4e4b\u95f4\u7684\u63a2\u7d22\u5e76\u63d0\u5347 MoE \u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u5c3d\u7ba1\u7a00\u758f\u95e8\u63a7 G (x; \u0398) \u53ef\u5728\u4e0d\u589e\u52a0\u76f8\u5e94\u8ba1\u7b97\u6210\u672c\u7684\u524d\u63d0\u4e0b\u663e\u8457\u6269\u5c55\u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7a7a\u95f4\uff0c\u4f46\u4e5f\u4f1a\u5bfc\u81f4\u8d1f\u8f7d\u5e73\u8861\u95ee\u9898\u3002\u8d1f\u8f7d\u5e73\u8861\u95ee\u9898\u662f\u6307\u5404\u4e13\u5bb6\u7684\u8d1f\u8f7d\u5206\u5e03\u4e0d\u5747 \u2014\u2014 \u67d0\u4e9b\u4e13\u5bb6\u88ab\u9891\u7e41\u4f7f\u7528\uff0c\u800c\u53e6\u4e00\u4e9b\u4e13\u5bb6\u5219\u5f88\u5c11\u88ab\u4f7f\u7528\u751a\u81f3\u5b8c\u5168\u4e0d\u4e0a\u573a\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6bcf\u4e2a MoE \u5c42\u90fd\u8981\u96c6\u6210\u4e00\u4e2a\u8f85\u52a9<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\uff0c\u5176\u4f5c\u7528\u662f\u6566\u4fc3\u6bcf<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u7684 token \u88ab\u5747\u5300\u5206\u914d\u7ed9\u5404\u4e2a\u4e13\u5bb6\u3002\u4ece\u6570\u5b66\u5f62\u5f0f\u63cf\u8ff0\u6765\u770b\uff0c\u9996\u5148\u5b9a\u4e49\u4e00\u4e2a\u5305\u542b T \u4e2a token \u7684<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark><mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark> B = {x_1 , x_2, &#8230; , x_\ud835\udc47 } \u4ee5\u53ca N \u4e2a\u4e13\u5bb6\u3002\u5219\u5bf9\u4e8e\u5176\u7684\u8f85\u52a9\u8d1f\u8f7d\u5e73\u8861\u635f\u5931\u5b9a\u4e49\u4e3a\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e89e057479cb1b715595a867c0ff66ea.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e89e057479cb1b715595a867c0ff66ea.png\"><\/a><\/p>\n<p>\u5176\u4e2d D_i \u662f\u5206\u914d\u7ed9\u4e13\u5bb6 i \u7684 token \u6bd4\u4f8b\uff0cP_i \u662f\u5206\u914d\u7ed9\u4e13\u5bb6 i \u7684\u95e8\u63a7\u6982\u7387\u6bd4\u4f8b\u3002\u4e3a\u4e86\u786e\u4fdd\u8be5<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u5728 N \u4e2a\u4e13\u5bb6\u4e4b\u95f4\u5747\u5300\u5206\u5e03\uff0c\u5e94\u5f53\u6700\u5c0f\u5316\u8d1f\u8f7d\u5e73\u8861<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark> L_{load-balancing}\u3002\u5f53\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u88ab\u5206\u914d\u4e86\u540c\u7b49\u6570\u91cf\u7684 token D_\ud835\udc56 = 1\/\ud835\udc41 \u548c\u540c\u7b49\u7684\u95e8\u63a7\u6982\u7387 P_\ud835\udc56 = 1\/\ud835\udc41 \u65f6\uff0c\u5373\u8fbe\u5230\u6700\u4f18\u6761\u4ef6\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-18edcd52d4cd9fc3e1c26bae72a9b073.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-18edcd52d4cd9fc3e1c26bae72a9b073.png\"><\/a><\/p>\n<p>\u6b64\u65f6\u5404\u4e13\u5bb6\u7684\u8d1f\u8f7d\u8fbe\u5230\u5e73\u8861\u3002<\/p>\n<p>\u5728\u540e\u6587\u4e2d\uff0c\u9664\u975e\u53e6\u6709\u660e\u786e\u8bf4\u660e\uff0c\u5219\u300cMoE\u300d\u8fd9\u4e00\u672f\u8bed\u5355\u6307\u300c\u7a00\u758f MoE\u300d\u3002<\/p>\n<p><strong>\u6df7\u5408\u4e13\u5bb6\u7684\u5206\u7c7b<\/strong><\/p>\n<p>\u4e3a\u4e86\u5e2e\u52a9\u7814\u7a76\u8005\u5728\u5927\u91cf\u91c7\u7528 MoE \u7684 LLM \u7814\u7a76\u4e2d\u627e\u5230\u76ee\u6807\uff0c\u8be5\u56e2\u961f\u5f00\u53d1\u4e86\u4e00\u5957\u5206\u7c7b\u65b9\u6cd5\uff0c\u6839\u636e\u4e09\u4e2a\u65b9\u9762\u5bf9\u8fd9\u4e9b\u6a21\u578b\u8fdb\u884c\u4e86\u5206\u7c7b\uff1a\u7b97\u6cd5\u8bbe\u8ba1\u3001\u7cfb\u7edf\u8bbe\u8ba1\u548c\u5e94\u7528\u3002<\/p>\n<p>\u56fe 3 \u5c55\u793a\u4e86\u8fd9\u79cd\u5206\u7c7b\u6cd5\u4ee5\u53ca\u4e00\u4e9b\u4ee3\u8868\u6027\u7814\u7a76\u6210\u679c\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-462d74bee0a1cec8e940aa1a5f6f7b85.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-462d74bee0a1cec8e940aa1a5f6f7b85.png\"><\/a><\/p>\n<p>\u4e0b\u9762\u5c06\u5168\u9762\u6df1\u5165\u5730\u4ecb\u7ecd\u5404\u7c7b\u522b\u7684\u60c5\u51b5\u3002<\/p>\n<p><strong>\u6df7\u5408\u4e13\u5bb6\u7684\u7b97\u6cd5\u8bbe\u8ba1<\/strong><\/p>\n<p><strong>\u95e8\u63a7\u51fd\u6570<\/strong><\/p>\n<p>\u95e8\u63a7\u51fd\u6570\uff08\u4e5f\u88ab\u79f0\u4e3a\u8def\u7531\u51fd\u6570\u6216\u8def\u7531\u5668\uff09\u662f\u6240\u6709 MoE \u67b6\u6784\u7684\u57fa\u7840\u7ec4\u4ef6\uff0c\u5176\u4f5c\u7528\u662f\u534f\u8c03\u4f7f\u7528\u4e13\u5bb6\u8ba1\u7b97\u4ee5\u53ca\u7ec4\u5408\u5404\u4e13\u5bb6\u7684\u8f93\u51fa\u3002<\/p>\n<p>\u6839\u636e\u5bf9\u6bcf\u4e2a\u8f93\u5165\u7684\u5904\u7406\u65b9\u6cd5\uff0c\u8be5\u95e8\u63a7\u53ef\u5206\u4e3a\u4e09\u79cd\u7c7b\u578b\uff1a\u7a00\u758f\u5f0f\u3001\u5bc6\u96c6\u5f0f\u548c soft \u5f0f\u3002\u5176\u4e2d\u7a00\u758f\u5f0f\u95e8\u63a7\u673a\u5236\u662f\u6fc0\u6d3b\u90e8\u5206\u4e13\u5bb6\uff0c\u800c\u5bc6\u96c6\u5f0f\u662f\u6fc0\u6d3b\u6240\u6709\u4e13\u5bb6\uff0csoft \u5f0f\u5219\u5305\u62ec\u5b8c\u5168\u53ef\u5fae\u65b9\u6cd5\uff0c\u5305\u62ec\u8f93\u5165 token \u878d\u5408\u548c\u4e13\u5bb6\u878d\u5408\u3002\u56fe 4 \u5c55\u793a\u4e86 MoE \u6a21\u578b\u4e2d\u4f7f\u7528\u7684\u5404\u79cd\u95e8\u63a7\u51fd\u6570\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d1f63f57e339070af650569a16cc0a3b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d1f63f57e339070af650569a16cc0a3b.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u7a00\u758f\u5f0f<\/p>\n<\/li>\n<\/ul>\n<p>\u7a00\u758f\u95e8\u63a7\u51fd\u6570\u5728\u5904\u7406\u5404\u4e2a\u8f93\u5165 token \u65f6\u4f1a\u6fc0\u6d3b\u88ab\u9009\u4e2d\u7684\u90e8\u5206\u4e13\u5bb6\uff0c\u8fd9\u53ef\u88ab\u89c6\u4e3a\u4e00\u79cd\u5f62\u5f0f\u7684\u6761\u4ef6\u8ba1\u7b97\u3002<\/p>\n<p>\u95e8\u63a7\u51fd\u6570\u53ef\u4ee5\u5b9e\u73b0\u591a\u79cd\u5f62\u5f0f\u7684\u95e8\u63a7\u51b3\u7b56\uff0c\u6bd4\u5982\u4e8c\u5143\u51b3\u7b56\u3001\u7a00\u758f\u6216\u8fde\u7eed\u51b3\u7b56\u3001\u968f\u673a\u6216\u786e\u5b9a\u6027\u51b3\u7b56\uff1b\u5176\u5df2\u7ecf\u5f97\u5230\u4e86\u6df1\u5165\u7684\u7814\u7a76\uff0c\u53ef\u4f7f\u7528\u5404\u79cd\u5f62\u5f0f\u7684<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u548c\u53cd\u5411\u4f20\u64ad\u6765\u8bad\u7ec3\u3002<\/p>\n<p>Shazeer et al. \u7684\u7814\u7a76\u300aOutrageously large neural networks: The sparsely-gated mixture-of-experts layer\u300b\u5f00\u521b\u6027\u5730\u63d0\u51fa\u4e86\u4e00\u79cd\u4f7f\u7528\u8f85\u52a9\u8d1f\u8f7d\u5e73\u8861\u635f\u5931\u7684\u53ef\u5fae\u5206\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u5176\u4e2d\u53ef\u6839\u636e\u9009\u53d6\u6982\u7387\u5bf9\u4e13\u5bb6\u8ba1\u7b97\u7684\u8f93\u51fa\u8fdb\u884c\u52a0\u6743\u3002\u8fd9\u4e3a\u95e8\u63a7\u8fc7\u7a0b\u5f15\u5165\u4e86\u53ef\u5fae\u6027\uff0c\u7531\u6b64\u53ef\u901a\u8fc7\u68af\u5ea6\u6765\u5f15\u5bfc\u95e8\u63a7\u51fd\u6570\u7684\u4f18\u5316\u3002<\/p>\n<p>\u540e\u6765\uff0c\u8fd9\u4e00\u8303\u5f0f\u4fbf\u6210\u4e86 MoE \u7814\u7a76\u9886\u57df\u7684\u4e3b\u5bfc\u8303\u5f0f\u3002\u7531\u4e8e\u8fd9\u79cd\u65b9\u6cd5\u4f1a\u9488\u5bf9\u6bcf\u4e2a\u8f93\u5165 token \u9009\u62e9\u4e13\u5bb6\uff0c\u56e0\u6b64\u53ef\u5c06\u5176\u770b\u4f5c\u662f token \u9009\u62e9\u5f0f\u95e8\u63a7\u51fd\u6570\u3002<\/p>\n<p>\u4ee5\u4e0b\u4e3a\u8fd9\u4e00\u5c0f\u8282\u7684\u8981\u70b9\uff0c\u8be6\u89c1\u539f\u8bba\u6587\uff1a<\/p>\n<ul>\n<li>\n<p>token \u9009\u62e9\u5f0f\u95e8\u63a7<\/p>\n<\/li>\n<li>\n<p>\u7528\u4e8e token \u9009\u62e9\u5f0f\u95e8\u63a7\u7684\u8f85\u52a9\u635f\u5931<\/p>\n<\/li>\n<li>\n<p>token \u9009\u62e9\u5f0f\u95e8\u63a7\u7684\u4e13\u5bb6\u5bb9\u91cf<\/p>\n<\/li>\n<li>\n<p>token \u9009\u62e9\u5f0f\u95e8\u63a7\u7684\u5176\u5b83\u8fdb\u5c55<\/p>\n<\/li>\n<li>\n<p>\u4e0d\u53ef\u8bad\u7ec3\u7684 token \u9009\u62e9\u5f0f\u95e8\u63a7<\/p>\n<\/li>\n<li>\n<p>\u4e13\u5bb6\u9009\u62e9\u5f0f\u95e8\u63a7<\/p>\n<\/li>\n<\/ul>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-183102a51742b9f27044fedd930b28fa.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-183102a51742b9f27044fedd930b28fa.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u5bc6\u96c6\u5f0f<\/p>\n<\/li>\n<\/ul>\n<p>\u5bc6\u96c6 MoE \u662f\u6307\u5904\u7406\u6bcf\u4e2a\u8f93\u5165\u65f6\u90fd\u6fc0\u6d3b\u6240\u6709\u4e13\u5bb6\u3002<\/p>\n<p>\u867d\u7136\u7a00\u758f MoE \u6709\u6548\u7387\u65b9\u9762\u7684\u4f18\u52bf\uff0c\u4f46\u5bc6\u96c6 MoE \u65b9\u5411\u4f9d\u7136\u5728\u4e0d\u65ad\u8fce\u6765\u521b\u65b0\u3002\u5c24\u5176\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u5bc6\u96c6\u6fc0\u6d3b\u5728 LoRA-MoE \u5fae\u8c03\u65b9\u9762\u8868\u73b0\u5f88\u597d\uff0c\u5e76\u4e14 LoRA \u4e13\u5bb6\u7684\u8ba1\u7b97\u5f00\u9500\u76f8\u5bf9\u8f83\u4f4e\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u6709\u6548\u7075\u6d3b\u5730\u96c6\u6210\u591a\u4e2a LoRA \u4ee5\u5b8c\u6210\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\u3002\u8fd9\u80fd\u4fdd\u7559\u539f\u59cb\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b\uff0c\u540c\u65f6\u4fdd\u7559\u5404\u4e2a LoRA \u9488\u5bf9\u5404\u4e2a\u4efb\u52a1\u7684\u72ec\u6709\u7279\u6027\u3002<\/p>\n<ul>\n<li>\n<p>soft \u5f0f<\/p>\n<\/li>\n<\/ul>\n<p>\u5bf9\u7a00\u758f MoE \u6765\u8bf4\uff0c\u4e00\u5927\u57fa\u672c\u79bb\u6563\u4f18\u5316\u96be\u9898\u662f\u5982\u4f55\u51b3\u5b9a\u4e3a\u6bcf\u4e2a token \u5206\u914d\u54ea\u4e9b\u5408\u9002\u7684\u4e13\u5bb6\u3002\u4e3a\u4e86\u786e\u4fdd\u4e13\u5bb6\u80fd\u5e73\u8861\u5730\u53c2\u4e0e\u5e76\u5c3d\u53ef\u80fd\u51cf\u5c11\u65e0\u5206\u914d token\uff0c\u8fd9\u901a\u5e38\u5fc5\u987b\u542f\u53d1\u5f0f\u7684\u8f85\u52a9\u635f\u5931\u3002\u5728\u6d89\u53ca\u5206\u5e03\u5916\u6570\u636e\u7684\u573a\u666f\uff08\u6bd4\u5982\u63a8\u7406<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u5c0f\u3001\u6709\u5168\u65b0\u8f93\u5165\u6216<mark data-type=\"tech_methods\" data-id=\"a5deb948-06e3-4875-acdb-35c268734006\">\u8fc1\u79fb\u5b66\u4e60<\/mark>\uff09\u4e2d\uff0c\u8fd9\u4e2a\u95ee\u9898\u5c24\u5176\u663e\u8457\u3002<\/p>\n<p>\u7c7b\u4f3c\u4e8e\u5bc6\u96c6 MoE\uff0csoft MoE \u65b9\u6cd5\u5728\u5904\u7406\u6bcf\u4e2a\u8f93\u5165\u65f6\u4e5f\u4f1a\u4f7f\u7528\u6240\u6709\u4e13\u5bb6\uff0c\u4ece\u800c\u7ef4\u6301\u5b8c\u5168\u53ef\u5fae\u6027\uff0c\u8fdb\u800c\u907f\u514d\u79bb\u6563\u4e13\u5bb6\u9009\u62e9\u65b9\u6cd5\u7684\u56fa\u6709\u95ee\u9898\u3002soft MoE \u4e0e\u5bc6\u96c6 MoE \u7684\u4e0d\u540c\u5728\u4e8e\u524d\u8005\u4f1a\u901a\u8fc7\u5bf9\u8f93\u5165 token \u6216\u4e13\u5bb6\u8fdb\u884c\u95e8\u63a7\u52a0\u6743\u7684\u878d\u5408\u6765\u7f13\u89e3\u8ba1\u7b97\u9700\u6c42\u3002<\/p>\n<p><strong>\u4e13\u5bb6<\/strong><\/p>\n<p>\u8fd9\u4e00\u8282\u4f1a\u4ecb\u7ecd MoE \u6846\u67b6\u5185<mark data-type=\"tech_methods\" data-id=\"9fbe8681-737e-4466-9126-9e37fc194ec6\">\u4e13\u5bb6\u7f51\u7edc<\/mark>\u7684\u67b6\u6784\uff0c\u5e76\u4f1a\u8ba8\u8bba\u534f\u8c03\u8fd9\u4e9b\u4e13\u5bb6\u7684\u6fc0\u6d3b\u7684\u95e8\u63a7\u51fd\u6570\u3002<\/p>\n<ul>\n<li>\n<p>\u7f51\u7edc\u7c7b\u578b<\/p>\n<\/li>\n<\/ul>\n<p>\u81ea\u4ece MoE \u88ab\u6574\u5408\u5230 Transformer \u67b6\u6784\u4e2d\u4ee5\u6765\uff0c\u5176\u901a\u5e38\u4f1a\u66ff\u4ee3\u8fd9\u4e9b\u6a21\u578b\u4e2d\u7684\u524d\u5411\u7f51\u7edc\uff08FFN\uff09\u6a21\u5757\u3002\u901a\u5e38\u6765\u8bf4\uff0cMoE \u5c42\u4e2d\u7684\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u4f1a\u590d\u5236\u5176\u66ff\u6362\u7684 FFN \u7684\u67b6\u6784\u3002<\/p>\n<p>\u8fd9\u79cd\u5c06 FFN \u7528\u4f5c\u4e13\u5bb6\u7684\u8303\u5f0f\u5230\u73b0\u5728\u4f9d\u7136\u662f\u4e3b\u6d41\uff0c\u4f46\u4eba\u4eec\u4e5f\u5bf9\u6b64\u505a\u4e86\u4e0d\u5c11\u6539\u8fdb\u3002<\/p>\n<ul>\n<li>\n<p>\u8d85<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark><\/p>\n<\/li>\n<\/ul>\n<p>\u7a00\u758f MoE \u6a21\u578b\u7684\u89c4\u6a21\u7531\u51e0\u4e2a\u5173\u952e\u8d85<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u63a7\u5236\uff0c\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>\u6bcf\u4e2a MoE \u5c42\u7684\u4e13\u5bb6\u6570\u91cf<\/p>\n<\/li>\n<li>\n<p>\u6bcf\u4e2a\u4e13\u5bb6\u7684\u5927\u5c0f<\/p>\n<\/li>\n<li>\n<p>MoE \u5c42\u5728\u6574\u4e2a\u6a21\u578b\u4e2d\u7684\u653e\u7f6e\u9891\u7387<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u8d85<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684\u9009\u62e9\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u4f1a\u6df1\u523b\u5f71\u54cd\u6a21\u578b\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u548c\u8ba1\u7b97\u6548\u7387\u3002\u56e0\u6b64\uff0c\u8981\u6839\u636e\u7279\u5b9a\u7684\u5e94\u7528\u8981\u6c42\u548c\u8ba1\u7b97\u57fa\u7840\u8bbe\u65bd\u6765\u9009\u62e9\u6700\u4f73\u8d85<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3002\u8868 2 \u7ed9\u51fa\u4e86\u4e00\u4e9b\u4f7f\u7528 MoE \u7684\u6a21\u578b\u7684\u914d\u7f6e\u60c5\u51b5\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-72eac861927f91f0ac99cb2f34a26826.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-72eac861927f91f0ac99cb2f34a26826.png\"><\/a><\/p>\n<p>\u53e6\u5916\uff0c\u8868 3 \u5217\u4e3e\u4e86\u4e00\u4e9b\u8fd1\u671f\u7684\u5f00\u6e90\u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6570\u91cf\u548c<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6027\u80fd\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1c6a7467f38661de106ae2906c04e71d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1c6a7467f38661de106ae2906c04e71d.png\"><\/a><\/p>\n<ul>\n<li>\n<p><mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark><\/p>\n<\/li>\n<\/ul>\n<p>\u57fa\u4e8e\u5bc6\u96c6 Transformer \u67b6\u6784\u6784\u5efa\u7684\u7a00\u758f MoE \u6a21\u578b\u91c7\u7528\u4e86\u4e0e BERT\u3001T5\u3001GPT \u548c LLAMA \u7b49\u9886\u5148\u7684\u5bc6\u96c6 LLM \u7c7b\u4f3c\u7684<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u3002<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u5df2\u7ecf\u4ece ReLU \u53d1\u5c55\u51fa\u4e86 GeLU\u3001GeGLU\u3001SwiGLU \u7b49\u66f4\u5148\u8fdb\u7684\u9009\u62e9\u3002<\/p>\n<p>\u8fd9\u4e00\u8d8b\u52bf\u4e5f\u6269\u5c55\u5230\u4e86 MoE \u6a21\u578b\u7684\u5176\u5b83\u7ec4\u4ef6\uff0c\u5b83\u4eec\u7ecf\u5e38\u6574\u5408\u5747\u65b9\u6839<mark data-type=\"tech_methods\" data-id=\"edceaca3-7a3d-42b3-9189-d80b8b99a6ad\">\u5c42\u5f52\u4e00\u5316<\/mark>\uff08RMSNorm\uff09\u3001\u5206\u7ec4<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u6ce8\u610f\u529b\uff08GQA\uff09\u548c\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165\uff08RoPE\uff09\u7b49\u6280\u672f\u3002<\/p>\n<ul>\n<li>\n<p>\u5171\u4eab\u4e13\u5bb6<\/p>\n<\/li>\n<\/ul>\n<p>DeepSpeed-MoE \u521b\u65b0\u6027\u5730\u5f15\u5165\u4e86\u6b8b\u5dee MoE\uff08Residual-MoE\uff09\u67b6\u6784\uff0c\u5176\u4e2d\u6bcf\u4e2a token \u90fd\u7531\u4e00\u4e2a\u56fa\u5b9a\u4e13\u5bb6\u5916\u52a0\u4e00\u4e2a\u95e8\u63a7\u9009\u62e9\u7684\u4e13\u5bb6\u8fdb\u884c\u5904\u7406\uff0c\u5b9e\u73b0\u4e86\u6bcf\u4e00\u5c42\u90fd\u6709\u4e24\u4e2a\u4e13\u5bb6\u53c2\u4e0e\u5904\u7406\uff0c\u540c\u65f6\u4e5f\u4e0d\u4f1a\u8ba9\u901a\u4fe1\u6210\u672c\u8d85\u8fc7 top-1 \u95e8\u63a7\u65b9\u6cd5\u3002\u8fd9\u79cd\u65b9\u6cd5\u662f\u628a\u95e8\u63a7\u9009\u62e9\u7684 MoE \u4e13\u5bb6\u5f53\u4f5c\u662f\u56fa\u5b9a\u5bc6\u96c6 FFN \u7684\u7ea0\u9519\u8f85\u52a9\u3002<\/p>\n<p>NLLB \u4e2d\u4f7f\u7528\u7684\u6761\u4ef6\u5f0f MoE \u8def\u7531\uff08CMR\/Conditional MoE Routing\uff09\u4e5f\u91c7\u7528\u4e86\u7c7b\u4f3c\u7684\u65b9\u6cd5\uff0c\u5c06\u5bc6\u96c6 FFN \u548c MoE \u5c42\u7684\u8f93\u51fa\u7ec4\u5408\u8d77\u6765\u4f7f\u7528\u3002<\/p>\n<p>\u5c06\u56fa\u5b9a FFN \u548c\u7a00\u758f MoE \u6574\u5408\u8d77\u6765\u7684\u8303\u5f0f\u901a\u5e38\u88ab\u79f0\u4e3a\u5171\u4eab\u4e13\u5bb6\uff0c\u5982\u56fe 5b \u6240\u793a\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a29058e80e1ea91cbdfb75da92371b92.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a29058e80e1ea91cbdfb75da92371b92.png\"><\/a><\/p>\n<p>\u8fd1\u671f\u6709 DeepSeekMoE\u3001OpenMoE\u3001Qwen1.5-MoE \u548c MoCLE \u7b49\u6a21\u578b\u91c7\u7528\u8fd9\u4e00\u8303\u5f0f\uff0c\u8868\u660e\u5176\u6b63\u5728\u6210\u4e3a\u4e00\u79cd\u4e3b\u6d41\u914d\u7f6e\u3002\u4e0d\u8fc7 DeepSeekMoE \u548c Qwen1.5-MoE \u91c7\u7528\u4e86\u591a\u4e2a\u5171\u4eab\u4e13\u5bb6\uff0c\u800c\u4e0d\u662f\u5355\u4e2a\u3002<\/p>\n<p><strong>\u6df7\u5408<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u9ad8\u6548\u578b\u4e13\u5bb6<\/strong><\/p>\n<p><mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u9ad8\u6548\u578b\u5fae\u8c03\uff08PEFT\uff09\u662f\u4e00\u79cd\u63d0\u5347\u5fae\u8c03\u6548\u7387\u7684\u65b9\u6cd5\u3002\u7b80\u5355\u6765\u8bf4\uff0cPEFT \u5c31\u662f\u5728\u5fae\u8c03\u65f6\u4ec5\u66f4\u65b0\u57fa\u7840\u6a21\u578b\u7684\u4e00\u5c0f\u90e8\u5206<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3002<\/p>\n<p>PEFT \u5f88\u6210\u529f\uff0c\u4f46\u7531\u4e8e\u5176\u53ef\u8bad\u7ec3\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u6709\u9650\u4ee5\u53ca\u53ef\u80fd\u5b58\u5728\u7684\u707e\u96be\u6027\u9057\u5fd8\u95ee\u9898\uff0c\u8be5\u65b9\u6cd5\u96be\u4ee5\u7528\u4e8e\u9700\u8981\u6cdb\u5316\u5230\u591a\u4e2a\u4efb\u52a1\u7684\u60c5\u51b5\u3002<\/p>\n<p>\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e9b\u5c40\u9650\uff0c\u6df7\u5408<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u9ad8\u6548\u578b\u4e13\u5bb6\uff08MoPE\uff09\u8bde\u751f\u4e86\uff0c\u5176\u5c06 MoE \u6846\u67b6\u4e0e PEFT \u6574\u5408\u5230\u4e86\u4e00\u8d77\u3002MoPE \u96c6\u6210\u4e86 MoE \u7684\u95e8\u63a7\u673a\u5236\u4e0e\u591a\u4e13\u5bb6\u67b6\u6784\uff0c\u540c\u65f6\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u4f7f\u7528\u4e86 PEFT \u6280\u672f\u8fdb\u884c\u6784\u5efa\u3002\u8fd9\u79cd\u5de7\u5999\u7684\u7ec4\u5408\u80fd\u6781\u5927\u63d0\u5347 PEFT \u5728\u591a\u4efb\u52a1\u573a\u666f\u4e2d\u7684\u6027\u80fd\u3002\u6b64\u5916\uff0c\u7531\u4e8e\u4f7f\u7528\u4e86 PEFT \u6765\u6784\u5efa\u4e13\u5bb6\uff0c\u56e0\u6b64 MoPE \u4f7f\u7528\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u4e5f\u66f4\u5c11\uff0c\u8d44\u6e90\u6548\u7387\u6bd4\u4f20\u7edf MoE \u6a21\u578b\u9ad8\u5f97\u591a\u3002<\/p>\n<p>MoPE \u878d\u5408\u4e86 MoE \u7684\u591a\u4efb\u52a1\u7279\u6027\u4e0e PEFT \u7684\u8d44\u6e90\u6548\u7387\uff0c\u662f\u4e00\u4e2a\u6781\u5177\u524d\u666f\u7684\u7814\u7a76\u65b9\u5411\u3002\u56fe 6 \u6839\u636e\u5728 Transformer \u6a21\u578b\u67b6\u6784\u4e2d\u7684\u4f4d\u7f6e\u5bf9 MoPE \u8fdb\u884c\u4e86\u5206\u7c7b\u3002\u81f3\u4e8e MoPE \u65b9\u9762\u66f4\u8be6\u7ec6\u7684\u7814\u7a76\u6210\u679c\u4ecb\u7ecd\uff0c\u8bf7\u53c2\u770b\u539f\u8bba\u6587\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8002c6f35647e2340c3f46e029f3bc52.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8002c6f35647e2340c3f46e029f3bc52.png\"><\/a><\/p>\n<p><strong>\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u6848<\/strong><\/p>\n<p>\u6df7\u5408\u4e13\u5bb6\u5728\u8fdb\u6b65\u53d1\u5c55\uff0c\u76f8\u5173\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u6848\u4e5f\u5728\u8fdb\u6b65\u53d1\u5c55\u3002<\/p>\n<p>\u521d\u59cb\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u6848\u9700\u8981\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3 MoE \u6a21\u578b\uff0c\u76f4\u63a5\u91c7\u7528\u8bad\u7ec3\u7684\u6a21\u578b\u914d\u7f6e\u6765\u6267\u884c\u63a8\u7406\u3002<\/p>\n<p>\u4f46\u73b0\u5728\uff0cMoE \u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u9762\u5df2\u7ecf\u51fa\u73b0\u4e86\u8bb8\u591a\u65b0\u8303\u5f0f\uff0c\u5305\u62ec\u7ec4\u5408\u5bc6\u96c6\u6a21\u578b\u548c\u7a00\u758f\u6a21\u578b\u7684\u4f18\u52bf\u5b9e\u73b0\u53d6\u957f\u8865\u77ed\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-64f369b3c0b6f2c74d29c0d13d4db497.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-64f369b3c0b6f2c74d29c0d13d4db497.png\"><\/a><\/p>\n<p>\u56fe 7 \u5c55\u793a\u4e86\u4e0e MoE \u76f8\u5173\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u6848\uff0c\u53ef\u4ee5\u770b\u5230\u65b0\u51fa\u73b0\u7684\u65b9\u6848\u53ef\u5206\u4e3a\u4e09\u7c7b\uff1a<\/p>\n<ul>\n<li>\n<p>\u5bc6\u96c6\u5230\u7a00\u758f\uff1a\u4ece\u5bc6\u96c6\u6a21\u578b\u8bad\u7ec3\u5f00\u59cb\uff0c\u9010\u6b65\u8fc7\u6e21\u5230\u7a00\u758f MoE \u914d\u7f6e\uff1b<\/p>\n<\/li>\n<li>\n<p>\u7a00\u758f\u5230\u5bc6\u96c6\uff1a\u6d89\u53ca\u5230\u5c06\u7a00\u758f MoE \u6a21\u578b\u964d\u683c\u4e3a\u5bc6\u96c6\u5f62\u5f0f\uff0c\u8fd9\u6709\u5229\u4e8e\u5c06\u63a8\u7406\u5b9e\u73b0\u4e3a\u786c\u4ef6\u5f62\u5f0f\uff1b<\/p>\n<\/li>\n<li>\n<p>\u4e13\u5bb6\u6a21\u578b\u878d\u5408\uff1a\u5c06\u591a\u4e2a\u9884\u8bad\u7ec3\u5bc6\u96c6\u4e13\u5bb6\u6a21\u578b\u6574\u5408\u6210\u4e00\u4e2a\u7edf\u4e00\u7684 MoE \u6a21\u578b\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>MoE \u7684\u884d\u751f\u6280\u672f<\/strong><\/p>\n<p>\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u542f\u53d1\u4e86\u8bb8\u591a\u4e0d\u540c\u7684\u53d8\u4f53\u6280\u672f\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0cXue et al. \u7684\u8bba\u6587\u300aGo wider instead of deeper\u300b\u63d0\u51fa\u4e86\u6a21\u578b\u5bbd\u5ea6\u589e\u5927\u7684 WideNet\uff0c\u5176\u505a\u6cd5\u662f\u5c06\u524d\u5411\u7f51\u7edc\uff08FFN\uff09\u66ff\u6362\u6210 MoE \u5c42\uff0c\u540c\u65f6\u7ef4\u6301 Transformer \u5c42\u4e0a\u7684\u5171\u4eab\u53ef\u8bad\u7ec3<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\uff0c\u4f46\u5f52\u4e00\u5316\u5c42\u9664\u5916\u3002<\/p>\n<p>\u53e6\u5916\u8fd8\u6709 Tan et al. \u63d0\u51fa\u7684 SYT\uff08\u7a00\u758f\u901a\u7528 Transformer\uff09\u3001Antoniak et al. \u63d0\u51fa\u7684 MoT\uff08\u6df7\u5408 token\uff09\u3001Choi et al. \u63d0\u51fa\u7684 SMoP\uff08\u7a00\u758f\u6df7\u5408\u63d0\u8bcd\uff09\u3001Chen et al. \u63d0\u51fa\u7684 Lifelong-MoE\u3001Raposo et al. \u63d0\u51fa\u7684 MoD\uff08\u6df7\u5408\u6df1\u5ea6\uff09\u7b49\u3002<\/p>\n<p>\u603b\u7ed3\u4e00\u4e0b\uff0cMoE \u884d\u751f\u6280\u672f\u7684\u53d1\u5c55\u63ed\u793a\u4e86\u4e00\u4e2a\u8d8b\u52bf\uff1aMoE \u7684\u529f\u80fd\u8d8a\u6765\u8d8a\u591a\uff0c\u8d8a\u6765\u8d8a\u80fd\u9002\u5e94\u4e0d\u540c\u7684\u9886\u57df\u3002<\/p>\n<p><strong>\u6df7\u5408\u4e13\u5bb6\u7684\u7cfb\u7edf\u8bbe\u8ba1<\/strong><\/p>\n<p>\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u867d\u7136\u80fd\u589e\u5f3a\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u80fd\u529b\uff0c\u4f46\u4e5f\u5e26\u6765\u4e86\u65b0\u7684\u6280\u672f\u6311\u6218\uff0c\u56e0\u4e3a\u5176\u5177\u6709\u7a00\u758f\u4e14\u52a8\u6001\u7684\u8ba1\u7b97\u8d1f\u8f7d\u3002<\/p>\n<p>GShard \u5f15\u5165\u4e86\u4e13\u5bb6\u5e76\u884c\u5316\uff08expert parallelism\uff09\uff0c\u53ef\u6839\u636e\u4e13\u5bb6\u80fd\u529b\u7684\u8d1f\u8f7d\u5e73\u8861\u9650\u5236\u6765<mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\u5207\u5206\u540e\u7684\u5c40\u90e8 token\uff0c\u4ece\u800c\u5b9e\u73b0\u5e76\u884c\u7684\u95e8\u63a7\u548c\u4e13\u5bb6\u8ba1\u7b97\u3002\u8be5\u8303\u5f0f\u5df2\u7ecf\u6210\u4e3a\u4fc3\u8fdb MoE \u6a21\u578b\u9ad8\u6548\u6269\u5c55\u7684\u57fa\u7840\u7b56\u7565\u3002\u6211\u4eec\u53ef\u4ee5\u5c06\u8be5\u65b9\u6cd5\u770b\u4f5c\u662f\u589e\u5f3a\u7248\u7684\u6570\u636e\u5e76\u884c\u5316 \u2014\u2014MoE \u5c42\u4e2d\u7684\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u88ab\u5206\u914d\u5230\u4e00\u53f0\u4e0d\u540c\u8bbe\u5907\uff0c\u540c\u65f6\u6240\u6709\u8bbe\u5907\u4e0a\u90fd\u91cd\u590d\u914d\u5907\u6240\u6709\u975e\u4e13\u5bb6\u5c42\u3002<\/p>\n<p>\u5982\u56fe 8a \u6240\u793a\uff0c\u4e13\u5bb6\u5e76\u884c\u5316\u7684\u5de5\u4f5c\u6d41\u7a0b\u662f\u6309\u987a\u5e8f\u6267\u884c\u4ee5\u4e0b\u64cd\u4f5c\uff1a\u95e8\u8def\u7531\u3001\u8f93\u5165\u7f16\u7801\u3001All-to-All <mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\u3001\u4e13\u5bb6\u8ba1\u7b97\u3001All-to-All \u7ec4\u5408\u3001\u8f93\u51fa\u89e3\u7801\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-10a42d64cba2ab36c979fc26edecd2c3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-10a42d64cba2ab36c979fc26edecd2c3.png\"><\/a><\/p>\n<p>\u4e00\u822c\u6765\u8bf4\uff0cGEMM \u7684\u8f93\u5165\u5927\u5c0f\u9700\u8981\u8db3\u591f\u5927\uff0c\u4ee5\u4fbf\u5145\u5206\u5229\u7528\u8ba1\u7b97\u8bbe\u5907\u3002\u56e0\u6b64\uff0c\u8981\u4f7f\u7528\u8f93\u5165\u7f16\u7801\u5c06\u540c\u4e00\u4e2a\u4e13\u5bb6\u7684\u8f93\u5165 token \u805a\u5408\u5230\u8fde\u7eed\u7684\u5185\u5b58\u7a7a\u95f4\u4e2d\uff0c\u8fd9\u7531\u95e8\u8def\u7531\u4e2d\u7684\u300ctoken &#8211; \u4e13\u5bb6<mark data-type=\"concepts\" data-id=\"8ec6a68f-ad96-4b85-ab72-6f8931886922\">\u6620\u5c04<\/mark>\u300d\u51b3\u5b9a\u3002\u4e4b\u540e\uff0cAll-to-All <mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\u7684\u4f5c\u7528\u662f\u5c06\u8f93\u5165 token \u5206\u53d1\u7ed9\u5404\u8bbe\u5907\u4e0a\u5bf9\u5e94\u7684\u4e13\u5bb6\u3002\u4e4b\u540e\u662f\u4e13\u5bb6\u7684\u672c\u5730\u5316\u8ba1\u7b97\u3002\u8ba1\u7b97\u5b8c\u6210\u540e\u518d\u901a\u8fc7 All-to-All \u7ec4\u5408\u6c47\u603b\uff0c\u7136\u540e\u89e3\u7801\u8f93\u51fa\uff0c\u6839\u636e\u95e8\u63a7\u7d22\u5f15\u6062\u590d\u539f\u59cb\u6570\u636e\u7684\u5e03\u5c40\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4e5f\u6709\u7814\u7a76\u8005\u63a2\u7d22\u4e13\u5bb6\u5e76\u884c\u5316\u4e0e\u5176\u5b83\u5df2\u6709\u5e76\u884c\u7b56\u7565\uff08\u6bd4\u5982<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u3001\u7ba1\u9053\u5316\u3001\u5e8f\u5217\u5e76\u884c\u5316\uff09\u7684\u534f\u540c\uff0c\u4ee5\u63d0\u5347 MoE \u6a21\u578b\u5728\u5927\u89c4\u6a21\u5206\u5e03\u5f0f\u73af\u5883\u4e2d\u7684\u53ef\u6269\u5c55\u6027\u548c\u6548\u7387\u3002<\/p>\n<p>\u56fe 8 \u4e2d\u7ed9\u51fa\u4e86\u4e00\u4e9b\u6df7\u5408\u5e76\u884c\u5316\u793a\u4f8b\uff0c\u5305\u62ec (b) \u6570\u636e + \u4e13\u5bb6 + <mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u5e76\u884c\u5316\u3001(c) \u6570\u636e + \u4e13\u5bb6 + \u7ba1\u9053\u5e76\u884c\u5316\u3001(d) \u4e13\u5bb6 + <mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u5e76\u884c\u3002<\/p>\n<p>\u9700\u8981\u8ba4\u8bc6\u5230\uff0c\u8ba1\u7b97\u6548\u7387\u3001\u901a\u4fe1\u8d1f\u8f7d\u3001\u5185\u5b58\u5360\u7528\u4e4b\u95f4\u5b58\u5728\u590d\u6742\u7684\u76f8\u4e92\u4f5c\u7528\uff0c\u5206\u5e03\u5f0f\u5e76\u884c\u5316\u7b56\u7565\u7684\u9009\u62e9\u4f1a\u5bf9\u5176\u4ea7\u751f\u5f71\u54cd\uff0c\u5e76\u4e14\u4e5f\u4f1a\u88ab\u4e0d\u540c\u7684\u786c\u4ef6\u914d\u7f6e\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u5728\u90e8\u7f72\u7528\u4e8e\u5b9e\u9645\u5e94\u7528\u7684\u7b56\u7565\u65f6\uff0c\u5fc5\u987b\u7ec6\u81f4\u5730\u6743\u8861\u8003\u8651\u5e76\u9488\u5bf9\u5177\u4f53\u573a\u666f\u8fdb\u884c\u8c03\u6574\u3002<\/p>\n<p>\u4e4b\u540e\uff0c\u8be5\u56e2\u961f\u5206\u8ba1\u7b97\u3001\u901a\u4fe1\u548c\u5b58\u50a8\u4e09\u5927\u677f\u5757\u4ecb\u7ecd\u4e86 MoE \u6a21\u578b\u5f00\u53d1\u6240\u9762\u4e34\u7684\u7cfb\u7edf\u8bbe\u8ba1\u96be\u9898\u4ee5\u53ca\u89e3\u51b3\u8fd9\u4e9b\u96be\u9898\u7684\u7814\u7a76\u6210\u679c\uff0c\u8be6\u89c1\u539f\u8bba\u6587\u3002\u8868 4 \u7ed9\u51fa\u4e86\u5f00\u6e90 MoE \u6846\u67b6\u7684\u6982\u51b5\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b90935ca30ef340dd8793b6733ce2490.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b90935ca30ef340dd8793b6733ce2490.png\"><\/a><\/p>\n<p><strong>\u6df7\u5408\u4e13\u5bb6\u7684\u5e94\u7528<\/strong><\/p>\n<p>\u5728\u5f53\u524d Transformer \u4e3b\u5bfc\u7684\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u9886\u57df\uff0c\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u8303\u5f0f\u9887\u5177\u5438\u5f15\u529b\uff0c\u56e0\u4e3a\u5176\u80fd\u5728\u4e0d\u7ed9\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u5f15\u5165\u8fc7\u5927\u8ba1\u7b97\u9700\u6c42\u7684\u524d\u63d0\u4e0b\u5927\u5e45\u63d0\u5347\u6a21\u578b\u80fd\u529b\u3002\u8fd9\u7c7b\u6280\u672f\u80fd\u663e\u8457 LLM \u5728\u591a\u79cd\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\uff0c\u751a\u81f3\u9020\u5c31\u4e86\u4e00\u4e9b\u8d85\u8d8a\u4eba\u7c7b\u6c34\u5e73\u7684 AI \u5e94\u7528\u3002<\/p>\n<p>\u6709\u4f20\u8a00\u8bf4\u5f3a\u5927\u5982\u65af\u7684 GPT-4 \u53ef\u80fd\u4e5f\u91c7\u7528\u4e86\u67d0\u79cd MoE \u67b6\u6784 \u2014\u2014 \u7531 8 \u4e2a 2200 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684\u4e13\u5bb6\u6784\u6210\uff0c\u5728\u591a\u6837\u5316\u7684\u6570\u636e\u96c6\u548c\u4efb\u52a1\u4e0a\u5b8c\u6210\u4e86\u8bad\u7ec3\uff0c\u5e76\u4f7f\u7528\u4e86\u4e00\u79cd 16 \u6b21\u8fed\u4ee3\u7684\u63a8\u7406\u8fc7\u7a0b\u3002\u6709\u5173\u8be5\u4f20\u8a00\u7684\u66f4\u591a\u8be6\u60c5\u53ef\u53c2\u9605<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u62a5\u9053\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650883517&amp;idx=2&amp;sn=4b795725b3341dc60b282a19e31692de&amp;chksm=84e481c3b39308d5758cc635fef1ffd6cf56e57e345e82ddce683db42fd2b6a741747c35abdb&amp;scene=21#wechat_redirect\" target=\"_blank\">\u7ec8\u6781\u300c\u63ed\u79d8\u300d\uff1aGPT-4 \u6a21\u578b\u67b6\u6784\u3001\u8bad\u7ec3\u6210\u672c\u3001\u6570\u636e\u96c6\u4fe1\u606f\u90fd\u88ab\u6252\u51fa\u6765\u4e86<\/a>\u300b\u3002<\/p>\n<p>\u6240\u4ee5\uff0c\u6beb\u4e0d\u5947\u602a MoE \u5728<mark data-type=\"tech_tasks\" data-id=\"c8ff5114-6cbb-49ca-8a89-3ee2826be0b4\">\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/mark>\u3001<mark data-type=\"tech_tasks\" data-id=\"6e614199-9e49-450e-9078-61fb2b122da9\">\u8ba1\u7b97\u673a\u89c6\u89c9<\/mark>\u3001<mark data-type=\"tech_tasks\" data-id=\"6ca1ea2d-6bca-45b7-9c93-725d288739c3\">\u63a8\u8350\u7cfb\u7edf<\/mark>\u548c\u591a\u6a21\u6001\u5e94\u7528\u4e2d\u904d\u5730\u5f00\u82b1\u4e86\u3002<\/p>\n<p>\u8fd9\u4e9b\u5e94\u7528\u672c\u8d28\u4e0a\u5c31\u9700\u8981\u4f7f\u7528\u6761\u4ef6\u8ba1\u7b97\u6765\u5927\u5e45\u63d0\u5347\u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\uff0c\u4ee5\u6b64\u589e\u5f3a\u6a21\u578b\u5728\u56fa\u5b9a\u8ba1\u7b97\u6210\u672c\u4e0b\u7684\u6027\u80fd\uff0c\u6216\u901a\u8fc7\u95e8\u63a7\u673a\u5236\u5b9e\u73b0\u52a8\u6001\u4e13\u5bb6\u9009\u62e9\u6765\u5b9e\u73b0\u9ad8\u6548<mark data-type=\"tech_methods\" data-id=\"e240bcd1-bf0d-4628-b794-f07f3bea7083\">\u591a\u4efb\u52a1\u5b66\u4e60<\/mark>\u3002<\/p>\n<p>\u8be5\u56e2\u961f\u4e5f\u4ecb\u7ecd\u4e86\u8fd9\u4e9b\u4e0d\u540c\u9886\u57df\u7684\u4ee3\u8868\u6027 MoE \u5e94\u7528\uff0c\u53ef\u5e2e\u52a9\u8bfb\u8005\u7406\u89e3\u5982\u4f55\u5c06 MoE \u7528\u4e8e\u5177\u4f53\u4efb\u52a1\u3002\u8be6\u89c1\u539f\u8bba\u6587\u3002<\/p>\n<p><strong>\u6311\u6218\u4e0e\u673a\u9047<\/strong><\/p>\n<p>\u6df7\u5408\u4e13\u5bb6\uff0c\u529f\u80fd\u5f3a\u5927\uff0c\u964d\u4f4e\u6210\u672c\uff0c\u63d0\u5347\u6027\u80fd\u3002\u524d\u666f\u867d\u597d\uff0c\u4ecd\u6709\u6311\u6218\u3002<\/p>\n<p>\u8fd9\u4e00\u8282\u4e2d\uff0c\u8be5\u56e2\u961f\u68b3\u7406\u4e86 MoE \u76f8\u5173\u7684\u5173\u952e\u6027\u6311\u6218\uff0c\u5e76\u6307\u51fa\u4e86\u6709\u5e0c\u671b\u83b7\u5f97\u91cd\u8981\u6210\u679c\u7684\u672a\u6765\u7814\u7a76\u65b9\u5411\u3002\u4e0b\u9762\u7b80\u8981\u5217\u51fa\u4e86\u8fd9\u4e9b\u6311\u6218\u548c\u7814\u7a76\u65b9\u5411\uff0c\u66f4\u591a\u8be6\u60c5\u8bf7\u67e5\u770b\u539f\u8bba\u6587\u3002<\/p>\n<ul>\n<li>\n<p>\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u8d1f\u8f7d\u5e73\u8861<\/p>\n<\/li>\n<li>\n<p>\u53ef\u6269\u5c55\u6027\u548c\u901a\u4fe1\u5f00\u9500<\/p>\n<\/li>\n<li>\n<p>\u4e13\u5bb6\u7684\u4e13\u4e1a\u5316\u548c\u534f\u4f5c<\/p>\n<\/li>\n<li>\n<p>\u7a00\u758f\u6fc0\u6d3b\u548c\u8ba1\u7b97\u6548\u7387<\/p>\n<\/li>\n<li>\n<p>\u6cdb\u5316\u548c\u7a33\u5065\u6027<\/p>\n<\/li>\n<li>\n<p>\u53ef\u89e3\u91ca\u6027\u548c\u900f\u660e\u6027<\/p>\n<\/li>\n<li>\n<p>\u6700\u4f18\u7684\u4e13\u5bb6\u67b6\u6784<\/p>\n<\/li>\n<li>\n<p>\u4e0e\u73b0\u6709\u6846\u67b6\u6574\u5408<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u6269\u5c55\u9605\u8bfb\uff1aMoE \u76f8\u5173\u62a5\u9053<\/strong><\/p>\n<p><strong>\u57fa\u7840\uff1a<\/strong><\/p>\n<ul>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650858005&amp;idx=1&amp;sn=167bc30ab9335962ac6586c66a79bb0f&amp;chksm=84e5266bb392af7d84e0c56ce4146a098ef603be05a5fd0b2ea8ab6e6829a4bae9b454489508&amp;scene=21#wechat_redirect\" target=\"_blank\">30 \u5e74\u5386\u53f2\u56de\u987e\uff0cJeff Dean\uff1a\u6211\u4eec\u6574\u7406\u4e86\u4e00\u4efd\u300c\u7a00\u758f\u4e13\u5bb6\u6a21\u578b\u300d\u7814\u7a76\u7efc\u8ff0<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"11\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650900669&amp;idx=5&amp;sn=573c5435e483b05b0f2ba40f5a10dd0e&amp;chksm=84e44cc3b393c5d5e8c4b54ea4159107b49717f39c59978c96b635b7c3c484d00b8d68c830bd&amp;scene=21#wechat_redirect\" target=\"_blank\">\u4e3a\u4ec0\u4e48\u57fa\u4e8e MoE \u7684\u5927\u6a21\u578b\u66f4\u503c\u5f97\u5173\u6ce8\uff1f<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650904974&amp;idx=3&amp;sn=669f3b0013c2e335cef576b55563ee84&amp;chksm=84e45df0b393d4e600023cc414837b77acd873d35defb35474d3b398e4ac97f00a898934e18b&amp;scene=21#wechat_redirect\" target=\"_blank\">\u88ab OpenAI\u3001Mistral AI \u5e26\u706b\u7684 MoE \u662f\u600e\u4e48\u56de\u4e8b\uff1f\u4e00\u6587\u8d2f\u901a\u4e13\u5bb6\u6df7\u5408\u67b6\u6784\u90e8\u7f72<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650850109&amp;idx=4&amp;sn=2589b1d2213883147def562569bff9db&amp;chksm=84e50743b3928e550062de9aea48bd475c1c65610db455775c77967be8e00f0547f9526ac0f5&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5438\u5f15<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u5708\u773c\u7403\u7684 MoE\uff0c\u4f1a\u6210\u4e3a NLP \u4e0e CV \u7684\u672a\u6765\u5417\uff1f<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650907458&amp;idx=1&amp;sn=14d6a69dda179cece342d0d798bacec7&amp;chksm=84e4673cb393ee2a32c68d49934ee1474f63f0b3eeb3b8926f6437c278b4fc3ac79f085cc997&amp;scene=21#wechat_redirect\" target=\"_blank\">\u624b\u628a\u624b\u6559\u4f60\uff0c\u4ece\u96f6\u5f00\u59cb\u5b9e\u73b0\u4e00\u4e2a\u7a00\u758f\u6df7\u5408\u4e13\u5bb6\u67b6\u6784<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08MoE\uff09<\/a><\/p>\n<\/li>\n<\/ul>\n<p><strong>\u524d\u6cbf\uff1a<\/strong><\/p>\n<ul>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650925512&amp;idx=3&amp;sn=7a737e23afe4f1f14659fc07baa8c2b1&amp;chksm=84e42db6b393a4a0d3a1223f57dad525f450cad58386609337cded89f8215e2498ddf7113f6d&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5355\u4e00\u4f5c\u8005\u8bba\u6587\uff0c\u8c37\u6b4c\u63d0\u51fa\u767e\u4e07\u4e13\u5bb6 Mixture\uff0c\u8d85\u8d8a\u5bc6\u96c6\u524d\u9988\u3001\u7a00\u758f MoE<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650917888&amp;idx=4&amp;sn=90c657686a0d4b7363686ec9588d04dd&amp;chksm=84e4087eb393816826bd8a9f44fcad32d75822eaa731021c1893634ed387d31d7cedd1683e89&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5fae\u8f6f\u8ba9 MoE \u957f\u51fa\u591a\u4e2a\u5934\uff0c\u5927\u5e45\u63d0\u5347\u4e13\u5bb6\u6fc0\u6d3b\u7387<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650906368&amp;idx=3&amp;sn=de29015cd3874bf021926bafa358f5c5&amp;chksm=84e45b7eb393d2689b3fa22b063234a9550ae9ceffc9cf1771621390f24039d2f97928e8da0d&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5c06\u591a\u6a21\u6001\u5927\u6a21\u578b\u7a00\u758f\u5316\uff0c3B \u6a21\u578b MoE-LLaVA \u5ab2\u7f8e LLaVA-1.5-7B<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650905254&amp;idx=3&amp;sn=71a840e160797a8052a8d6ea4fcf4dec&amp;chksm=84e45ed8b393d7ce34f461ca0b195b51a26aad975272cccd18a2a2e7dc2a316a2f800eabec95&amp;scene=21#wechat_redirect\" target=\"_blank\">MoE \u4e0e Mamba \u5f3a\u5f3a\u8054\u5408\uff0c\u5c06\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u6269\u5c55\u5230\u6570\u767e\u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark><\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650912509&amp;idx=1&amp;sn=cab582849245ef9382317ec13f820dfc&amp;chksm=84e47283b393fb95bf5f8b793eecb5280c5a3f67dd14a097e7aaa947e9dd2e04a6fffab5b1a0&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5f00\u6e90\u5927\u6a21\u578b\u738b\u5ea7\u518d\u6613\u4e3b\uff0c1320 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark> DBRX \u4e0a\u7ebf\uff0c\u57fa\u7840\u3001\u5fae\u8c03\u6a21\u578b\u90fd\u6709<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650915526&amp;idx=5&amp;sn=78f3444380c9ca6e7146a881a7231215&amp;chksm=84e406b8b3938fae7f47b441828fff2d259251938f0ac8bab4df6e9b198688710b3ec8013033&amp;scene=21#wechat_redirect\" target=\"_blank\">CVPR 2024 | \u57fa\u4e8e MoE \u7684\u901a\u7528\u56fe\u50cf\u878d\u5408\u6a21\u578b\uff0c\u6dfb\u52a0 2.8% <mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5b8c\u6210\u591a\u9879\u4efb\u52a1<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650872496&amp;idx=4&amp;sn=de6400986f9d53f85bd70066b1f7a878&amp;chksm=84e4deceb39357d805a3456afb5e1ed78dc36a3752b27e2e6aee6e5696a1f789fd87d86f0e70&amp;scene=21#wechat_redirect\" target=\"_blank\">CVPR 2023 | \u6a21\u5757\u5316 MoE \u5c06\u6210\u4e3a\u89c6\u89c9<mark data-type=\"tech_methods\" data-id=\"e240bcd1-bf0d-4628-b794-f07f3bea7083\">\u591a\u4efb\u52a1\u5b66\u4e60<\/mark>\u57fa\u7840\u6a21\u578b<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650907658&amp;idx=2&amp;sn=e6dad13b58a1ea9298c81e537f022041&amp;chksm=84e46074b393e962eb67f32e0f56357678a211d9cdad63ac28174cdbffc896662c35ffcf22a7&amp;scene=21#wechat_redirect\" target=\"_blank\">\u8c37\u6b4c Gemini1.5 \u706b\u901f\u4e0a\u7ebf\uff1aMoE \u67b6\u6784\uff0c100 \u4e07\u4e0a\u4e0b\u6587<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650911073&amp;idx=1&amp;sn=a915a4e1c32154400adeadbeb853925a&amp;chksm=84e4751fb393fc09a702a547b344f7daa5a5d639679f80f35b7333cdf86a44d1901d04444f06&amp;scene=21#wechat_redirect\" target=\"_blank\">\u82f9\u679c\u5927\u6a21\u578b MM1 \u6740\u5165\u573a\uff1a300 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3001\u591a\u6a21\u6001\u3001MoE \u67b6\u6784\uff0c\u8d85\u534a\u6570\u4f5c\u8005\u662f\u534e\u4eba<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650902476&amp;idx=2&amp;sn=05a810a5308474855903090833005521&amp;chksm=84e44bb2b393c2a4ce0ad46ad7214c46d07ce1f17f9f1bb62aabddcfb70c8d5c2059774dca50&amp;scene=21#wechat_redirect\" target=\"_blank\">8x7B MoE \u4e0e Flash Attention 2 \u7ed3\u5408\uff0c\u4e0d\u5230 10 \u884c\u4ee3\u7801\u5b9e\u73b0\u5feb\u901f\u63a8\u7406<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650906843&amp;idx=4&amp;sn=8a4a960e42ad3e10d1805eaf7ee92de7&amp;chksm=84e464a5b393edb356d83f5bdcd46d613dec70fa599247bb3979cd6b5a0c196885bd2fdb8ffa&amp;scene=21#wechat_redirect\" target=\"_blank\">\u6253\u7834 MoE \u8bad\u7ec3\u6548\u7387\u4e0e\u6027\u80fd\u74f6\u9888\uff0c<mark data-type=\"institutions\" data-id=\"48d62db6-3e47-4e93-8b9d-b1283a7ea419\">\u534e\u4e3a<\/mark>\u76d8\u53e4\u7a00\u758f\u5927\u6a21\u578b\u5168\u65b0\u67b6\u6784 LocMoE \u51fa\u7089<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"11\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650920390&amp;idx=4&amp;sn=0bc7ce29e6bcff325da99316b20e377e&amp;chksm=84e411b8b39398ae368f15554e3db1a2b16037e264124274a5599c184614e237b360847c972b&amp;scene=21#wechat_redirect\" target=\"_blank\">\u5355\u4e2a 4090 \u53ef\u63a8\u7406\uff0c2000 \u4ebf\u7a00\u758f\u5927\u6a21\u578b\u300c\u5929\u5de5 MoE\u300d\u5f00\u6e90<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650913885&amp;idx=5&amp;sn=4c5362a45c3a460ac2c3038b4a947379&amp;chksm=84e47823b393f13505f538ca74a1e7dcacce8932efb86b38cf6f05fef81f7f04e5a3330ec403&amp;scene=21#wechat_redirect\" target=\"_blank\">Mistral \u5f00\u6e90 8X22B \u5927\u6a21\u578b\uff0cOpenAI \u66f4\u65b0 GPT-4 Turbo \u89c6\u89c9\uff0c\u90fd\u5728\u6b3a\u8d1f\u8c37\u6b4c<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650899930&amp;idx=1&amp;sn=7dfa496cd639127d6779e478b829b07c&amp;chksm=84e441a4b393c8b2a57030622e16874be29f1103bc0ca14d10f8408dec71e6d6988d06152958&amp;scene=21#wechat_redirect\" target=\"_blank\">\u4e00\u6761\u78c1\u529b\u94fe\u63a5\u5e2d\u5377 AI \u5708\uff0c87GB \u79cd\u5b50\u76f4\u63a5\u5f00\u6e90 8x7B MoE \u6a21\u578b<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"11\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650914415&amp;idx=4&amp;sn=838e9be75a17c589aae3e83ca1d28174&amp;chksm=84e47a11b393f30748039e2573824e65d5013bb864aa0863bd691898a109aae3f3f1fe8dd786&amp;scene=21#wechat_redirect\" target=\"_blank\">\u6bd4 MoE \u66f4\u6709\u6f5c\u529b\uff1f\u8fdb\u5316\u7b97\u6cd5\u878d\u5408\u6a21\u578b\u7684\u65b0\u8def\u5f84\u662f\u5426\u503c\u5f97\u4e00\u8bd5\uff1f<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650886540&amp;idx=5&amp;sn=b656504d2a1ac8d555a8c68ccc0f7ded&amp;chksm=84e495f2b3931ce4b8c1b3f75fde8929af3c1a1ca49e6e543fea4fcdfb70c5d0a6f33c4e6f74&amp;scene=21#wechat_redirect\" target=\"_blank\">\u6e05\u534e\u53d1\u5e03 SmartMoE\uff1a\u4e00\u952e\u5b9e\u73b0\u9ad8\u6027\u80fd MoE \u7a00\u758f\u5927\u6a21\u578b\u5206\u5e03\u5f0f\u8bad\u7ec3<\/a><\/p>\n<\/li>\n<li>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650917057&amp;idx=4&amp;sn=2d3225438117cdea15e417b6aa6b3d1e&amp;chksm=84e40cbfb39385a9f367174f0721a69e4cacb0eb88a7ccc960029abe75285eb677ac938f87d7&amp;scene=21#wechat_redirect\" target=\"_blank\">\u4e00\u5757\u94b1 100 \u4e07 token\uff0c\u8d85\u5f3a MoE \u6a21\u578b\u5f00\u6e90\uff0c\u6027\u80fd\u76f4\u903c GPT-4-Turbo<\/a><\/p>\n<\/li>\n<\/ul>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-26-5\" target=\"_blank\">\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a [&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":[71,73,68,55,67,66],"class_list":["post-34758","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-73","tag-68","tag-55","tag-67","tag-66"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09 - \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=34758\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\",\"datePublished\":\"2024-12-12T03:57:43+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758\"},\"wordCount\":385,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-aca632cad1906584d4142ac8ef49d687.png\",\"keywords\":[\"RAG\",\"\u591a\u6a21\u6001\",\"\u5927\u6a21\u578b\",\"\u673a\u5668\u5b66\u4e60\",\"\u81ea\u7136\u8bed\u8a00\u5904\u7406\",\"\u8ba1\u7b97\u673a\u89c6\u89c9\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34758#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=34758\",\"name\":\"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-aca632cad1906584d4142ac8ef49d687.png\",\"datePublished\":\"2024-12-12T03:57:43+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34758\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-aca632cad1906584d4142ac8ef49d687.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-aca632cad1906584d4142ac8ef49d687.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34758#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\"}]},{\"@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":"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09 - \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=34758","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=34758#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=34758"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09","datePublished":"2024-12-12T03:57:43+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=34758"},"wordCount":385,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=34758#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png","keywords":["RAG","\u591a\u6a21\u6001","\u5927\u6a21\u578b","\u673a\u5668\u5b66\u4e60","\u81ea\u7136\u8bed\u8a00\u5904\u7406","\u8ba1\u7b97\u673a\u89c6\u89c9"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=34758#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34758","url":"https:\/\/17aitech.com\/?p=34758","name":"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34758#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34758#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png","datePublished":"2024-12-12T03:57:43+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34758#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34758"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34758#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aca632cad1906584d4142ac8ef49d687.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34758#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u7b97\u6cd5\u3001\u7cfb\u7edf\u548c\u5e94\u7528\uff0c\u4e09\u4e2a\u89c6\u89d2\u5168\u9762\u8bfb\u61c2\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09"}]},{"@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\/34758","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=34758"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34758\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34758"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34758"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34758"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}