{"id":35037,"date":"2024-12-17T10:59:37","date_gmt":"2024-12-17T02:59:37","guid":{"rendered":"https:\/\/17aitech.com\/?p=35037"},"modified":"2024-12-17T10:59:37","modified_gmt":"2024-12-17T02:59:37","slug":"%e6%b7%b7%e5%90%88%e4%b8%93%e5%ae%b6%e6%9b%b4%e6%9c%89%e4%b8%bb%e8%a7%81%e4%ba%86%ef%bc%8c%e8%83%bd%e6%84%9f%e7%9f%a5%e5%a4%9a%e6%a8%a1%e6%80%81%e5%88%86%e6%83%85%e5%86%b5%e8%a1%8c%e4%ba%8b%ef%bc%8cme","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35037","title":{"rendered":"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-11-2\" target=\"_blank\">\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"13\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<p>\u6df7\u5408\u4e13\u5bb6\uff0c\u4e5f\u5f97\u672f\u4e1a\u6709\u4e13\u653b\u3002<\/p>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section>\u5bf9\u4e8e\u76ee\u524d\u7684\u6df7\u5408\u6a21\u6001\u57fa\u7840\u6a21\u578b\uff0c\u5e38\u7528\u7684\u67b6\u6784\u8bbe\u8ba1\u662f\u878d\u5408\u7279\u5b9a\u6a21\u6001\u7684\u7f16\u7801\u5668\u6216\u89e3\u7801\u5668\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u5b58\u5728\u5c40\u9650\uff1a\u65e0\u6cd5\u6574\u5408\u4e0d\u540c\u6a21\u6001\u7684\u4fe1\u606f\uff0c\u4e5f\u96be\u4ee5\u8f93\u51fa\u5305\u542b\u591a\u79cd\u6a21\u6001\u7684\u5185\u5bb9\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u514b\u670d\u8fd9\u4e00\u5c40\u9650\uff0cMeta FAIR \u7684 Chameleon \u56e2\u961f\u5728\u8fd1\u671f\u7684\u8bba\u6587\u300aChameleon: Mixed-modal early-fusion foundation models\u300b\u4e2d\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u5355\u4e00 Transformer \u67b6\u6784\uff0c\u5b83\u53ef\u4ee5\u6839\u636e\u4e0b\u4e00\u4e2a token \u7684\u9884\u6d4b\u76ee\u6807\uff0c\u5bf9\u7531\u79bb\u6563\u56fe\u50cf\u548c\u6587\u672c token \u7ec4\u6210\u7684\u6df7\u5408\u6a21\u6001\u5e8f\u5217\u8fdb\u884c\u5efa\u6a21\uff0c\u4ece\u800c\u5728\u4e0d\u540c\u6a21\u6001\u4e4b\u95f4\u8fdb\u884c\u65e0\u7f1d\u63a8\u7406\u548c\u751f\u6210\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg\"><\/a><\/p>\n<section>\u5728\u7ea6 10 \u4e07\u4ebf\u6df7\u5408\u6a21\u6001 token \u4e0a\u5b8c\u6210\u9884\u8bad\u7ec3\u4e4b\u540e\uff0cChameleon \u8868\u73b0\u51fa\u4e86\u9002\u5e94\u5e7f\u6cdb\u7684\u89c6\u89c9\u548c\u8bed\u8a00\u80fd\u529b\uff0c\u80fd\u5f88\u597d\u5730\u5904\u7406\u591a\u79cd\u4e0d\u540c\u7684\u4e0b\u6e38\u4efb\u52a1\u3002Chameleon \u5728\u751f\u6210\u6df7\u5408\u6a21\u6001\u957f\u56de\u7b54\u4efb\u52a1\u7684\u8868\u73b0\u5c24\u5176\u4eae\u773c\uff0c\u5b83\u751a\u81f3\u6253\u8d25\u4e86 Gemini 1.0 Pro \u548c GPT-4V \u7b49\u5546\u7528\u6a21\u578b\u3002\u7136\u800c\u5bf9\u4e8e Chameleon \u8fd9\u6837\u5404\u79cd\u6a21\u6001\u4f1a\u5728\u6a21\u578b\u8bad\u7ec3\u7684\u65e9\u671f\u6df7\u5408\u8d77\u6765\u7684\u6a21\u578b\uff0c\u60f3\u8981\u62d3\u5c55\u5b83\u7684\u80fd\u529b\uff0c\u9700\u8981\u6295\u5165\u5927\u91cf\u7b97\u529b\u3002<\/section>\n<section><\/section>\n<section>\u57fa\u4e8e\u4ee5\u4e0a\u95ee\u9898\uff0cMeta FAIR \u7684\u56e2\u961f\u5bf9\u8def\u7531\u5f0f\u7a00\u758f\u67b6\u6784\uff08routed sparse architecture\uff09\u8fdb\u884c\u4e86\u4e00\u756a\u7814\u7a76\u63a2\u7d22\uff0c\u63d0\u51fa\u4e86 MoMa\uff1a\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408\u67b6\u6784\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-5c23c51c089bab0d9cce14e867281989.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-5c23c51c089bab0d9cce14e867281989.png\"><\/a><\/p>\n<\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aMoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2407.21770<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4e4b\u524d\u5df2\u6709\u7814\u7a76\u8868\u660e\uff0c\u8fd9\u7c7b\u67b6\u6784\u53ef\u4ee5\u6709\u6548\u5730\u6269\u5c55\u5355\u6a21\u6001\u7684\u57fa\u7840\u6a21\u578b\u7684\u80fd\u529b\uff0c\u4e5f\u53ef\u4ee5\u589e\u5f3a\u591a\u6a21\u6001\u5bf9\u6bd4\u5b66\u4e60\u6a21\u578b\u7684\u6027\u80fd\u3002\u4f46\u662f\uff0c\u5c06\u5176\u7528\u4e8e\u8f83\u65e9\u5c06\u5404\u79cd\u6a21\u6001\u878d\u5408\u7684\u6a21\u578b\u8bad\u7ec3\u8fd8\u662f\u4e00\u4e2a\u673a\u9047\u4e0e\u6311\u6218\u5e76\u5b58\u7684\u8bfe\u9898\uff0c\u8fd8\u5c11\u6709\u4eba\u7814\u7a76\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u7684\u7814\u7a76\u57fa\u4e8e\u8fd9\u4e00\u6d1e\u89c1\uff1a\u4e0d\u540c\u6a21\u6001\u5177\u6709\u56fa\u6709\u7684\u5f02\u6784\u6027 \u2014\u2014 \u6587\u672c\u548c\u56fe\u50cf token \u5177\u6709\u4e0d\u540c\u7684\u4fe1\u606f\u5bc6\u5ea6\u548c\u5197\u4f59\u6a21\u5f0f\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5c06\u8fd9\u4e9b token \u6574\u5408\u6210\u7edf\u4e00\u7684\u878d\u5408\u67b6\u6784\u7684\u540c\u65f6\uff0c\u8be5\u56e2\u961f\u4e5f\u63d0\u51fa\u901a\u8fc7\u6574\u5408\u9488\u5bf9\u5177\u4f53\u6a21\u6001\u7684\u6a21\u5757\u6765\u8fdb\u4e00\u6b65\u4f18\u5316\u8be5\u6846\u67b6\u3002\u8be5\u56e2\u961f\u5c06\u8fd9\u4e00\u6982\u5ff5\u79f0\u4e3a\u6a21\u6001\u611f\u77e5\u578b\u7a00\u758f\u6027\uff08modality-aware sparsity\uff09\uff0c\u7b80\u79f0 MaS\uff1b\u5176\u80fd\u8ba9\u6a21\u578b\u66f4\u597d\u5730\u6355\u83b7\u6bcf\u4e2a\u6a21\u6001\u7684\u7279\u5f81\uff0c\u540c\u65f6\u8fd8\u80fd\u901a\u8fc7\u90e8\u5206\u53c2\u6570\u5171\u4eab\u548c\u6ce8\u610f\u529b\u673a\u5236\u7ef4\u6301\u5f3a\u5927\u7684\u8de8\u6a21\u6001\u6574\u5408\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u4e4b\u524d\u7684 VLMo\u3001BEiT-3 \u548c VL-MoE \u7b49\u7814\u7a76\u5df2\u7ecf\u91c7\u7528\u4e86\u6df7\u5408\u6a21\u6001\u4e13\u5bb6\uff08MoME\/mixture-of-modality-experts\uff09\u65b9\u6cd5\u6765\u8bad\u7ec3\u89c6\u89c9 &#8211; \u8bed\u8a00\u7f16\u7801\u5668\u548c\u63a9\u7801\u5f0f\u8bed\u8a00\u5efa\u6a21\uff0c\u6765\u81ea FAIR \u7684\u7814\u7a76\u56e2\u961f\u66f4\u8fdb\u4e00\u6b65\u5c06 MoE \u7684\u53ef\u7528\u8303\u56f4\u53c8\u63a8\u8fdb\u4e86\u4e00\u6b65\u3002<\/section>\n<section><\/section>\n<section><strong>\u6a21\u578b\u67b6\u6784<\/strong><\/section>\n<section><\/section>\n<section><strong>\u65e9\u671f\u878d\u5408<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6587\u63d0\u51fa\u7684\u65b0\u6a21\u578b\u57fa\u4e8e Chameleon \u7684\u65e9\u671f\u878d\u5408\u67b6\u6784\uff0c\u5176\u505a\u6cd5\u662f\u5728\u4e00\u4e2a\u7edf\u4e00 Transformer \u4e2d\uff0c\u5c06\u56fe\u50cf\u548c\u6587\u672c\u8868\u793a\u6210\u4e00\u7cfb\u5217\u79bb\u6563 token\u3002Chameleon \u7684\u6838\u5fc3\u662f\u4e00\u4e2a\u57fa\u4e8e Transformer \u7684\u6a21\u578b\uff0c\u5176\u4f1a\u5728\u56fe\u50cf\u548c\u6587\u672c token \u7684\u7ec4\u5408\u5e8f\u5217\u4e0a\u5e94\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3002\u8fd9\u80fd\u8ba9\u8be5\u6a21\u578b\u6355\u83b7\u6a21\u6001\u5185\u548c\u6a21\u6001\u95f4\u7684\u590d\u6742\u5173\u8054\u3002\u8be5\u6a21\u578b\u7684\u8bad\u7ec3\u4f7f\u7528\u7684\u76ee\u6807\u662f\u4e0b\u4e00 token \u9884\u6d4b\u76ee\u6807\uff0c\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u751f\u6210\u6587\u672c\u548c\u56fe\u50cf token\u3002<\/section>\n<section><\/section>\n<section>\u5728 Chameleon \u4e2d\uff0c\u56fe\u50cf\u7684 token \u5316\u65b9\u6848\u91c7\u7528\u4e86\u4e00\u4e2a\u5b66\u4e60\u578b\u56fe\u50cf\u5206\u8bcd\u5668\uff0c\u5b83\u5c06\u57fa\u4e8e\u5927\u5c0f\u4e3a 8192 \u7684 codebook \u5c06 512 \u00d7 512 \u7684\u56fe\u50cf\u7f16\u7801\u6210 1024 \u4e2a\u79bb\u6563 token\u3002\u5bf9\u4e8e\u6587\u672c\u7684\u5206\u8bcd\u5c06\u4f7f\u7528\u4e00\u4e2a\u8bcd\u8868\u5927\u5c0f\u4e3a 65,536 \u7684 BPE \u5206\u8bcd\u5668\uff0c\u5176\u4e2d\u5305\u542b\u56fe\u50cf token\u3002\u8fd9\u79cd\u7edf\u4e00\u7684\u5206\u8bcd\u65b9\u6cd5\u53ef\u4ee5\u8ba9\u6a21\u578b\u65e0\u7f1d\u5904\u7406\u56fe\u50cf\u548c\u6587\u672c token \u4ea4\u7ec7\u9519\u6742\u7684\u4efb\u610f\u5e8f\u5217\u3002<\/section>\n<section><\/section>\n<section>\u501f\u52a9\u8fd9\u79cd\u65b9\u6cd5\uff0c\u65b0\u6a21\u578b\u7ee7\u627f\u4e86\u8868\u5f81\u7edf\u4e00\u3001\u7075\u6d3b\u6027\u597d\u3001\u53ef\u6269\u5c55\u6027\u9ad8\u3001\u652f\u6301\u7aef\u5230\u7aef\u5b66\u4e60\u8fd9\u4e9b\u4f18\u70b9\u3002<\/section>\n<section><\/section>\n<section>\u5728\u6b64\u57fa\u7840\u4e0a\uff08\u56fe 1a\uff09\uff0c\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\u65e9\u878d\u5408\u6a21\u578b\u7684\u6548\u7387\u548c\u6027\u80fd\uff0c\u8be5\u56e2\u961f\u8fd8\u5f15\u5165\u4e86\u6a21\u6001\u611f\u77e5\u578b\u7a00\u758f\u6027\u6280\u672f\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-54bbc03f4c99819ac3e6c7736fd15056.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-54bbc03f4c99819ac3e6c7736fd15056.png\"><\/a><\/p>\n<\/section>\n<section><strong>\u5bbd\u5ea6\u6269\u5c55\uff1a\u6a21\u6001\u611f\u77e5\u578b\u6df7\u5408\u4e13\u5bb6<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u79cd\u5bbd\u5ea6\u6269\u5c55\u65b9\u6cd5\uff1a\u5c06\u6a21\u6001\u611f\u77e5\u578b\u6a21\u5757\u7a00\u758f\u6027\u96c6\u6210\u5230\u524d\u5411\u6a21\u5757\u4e2d\uff0c\u4ece\u800c\u6269\u5c55\u6807\u51c6\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u67b6\u6784\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u65b9\u6cd5\u57fa\u4e8e\u8fd9\u4e00\u6d1e\u89c1\uff1a\u4e0d\u540c\u6a21\u6001\u7684 token \u6709\u5404\u81ea\u4e0d\u540c\u7684\u7279\u5f81\u548c\u4fe1\u606f\u5bc6\u5ea6\u3002<\/section>\n<section><\/section>\n<section>\u901a\u8fc7\u4e3a\u6bcf\u4e2a\u6a21\u6001\u6784\u5efa\u4e0d\u540c\u7684\u4e13\u5bb6\u5206\u7ec4\uff0c\u53ef\u8ba9\u6a21\u578b\u5f00\u53d1\u51fa\u4e13\u95e8\u7684\u5904\u7406\u8def\u5f84\uff0c\u540c\u65f6\u7ef4\u6301\u8de8\u6a21\u6001\u7684\u4fe1\u606f\u6574\u5408\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section>\u56fe 1b \u5c55\u793a\u4e86\u8fd9\u79cd\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408\uff08MoMa\uff09\u7684\u5173\u952e\u7ec4\u4ef6\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u5148\u662f\u5bf9\u5404\u4e2a\u7279\u5b9a\u6a21\u6001\u7684\u4e13\u5bb6\u8fdb\u884c\u5206\u7ec4\uff0c\u7136\u540e\u5b9e\u73b0\u5206\u5c42\u8def\u7531\uff08\u5206\u4e3a\u6a21\u6001\u611f\u77e5\u578b\u8def\u7531\u548c\u6a21\u6001\u5185\u8def\u7531\uff09\uff0c\u6700\u540e\u9009\u62e9\u4e13\u5bb6\u3002\u8be6\u7ec6\u8fc7\u7a0b\u53c2\u89c1\u539f\u8bba\u6587\u3002<\/section>\n<section><\/section>\n<section>\u603b\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u8f93\u5165 token x\uff0cMoMa \u6a21\u5757\u7684\u5f62\u5f0f\u5316\u5b9a\u4e49\u4e3a\uff1a<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-b77446bec0f9fba9b899eae2fce70cba.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-b77446bec0f9fba9b899eae2fce70cba.png\"><\/a><\/p>\n<section>\u5728 MoMa \u8ba1\u7b97\u4e4b\u540e\uff0c\u8be5\u56e2\u961f\u53c8\u8fdb\u4e00\u6b65\u4f7f\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5\u548c\u00a0Swin Transformer \u5f52\u4e00\u5316\u3002<\/section>\n<section><\/section>\n<section><\/section>\n<section><strong>Mixture-of-Depths\uff08MoD\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u4e4b\u524d\u4e5f\u6709\u7814\u7a76\u8005\u63a2\u7d22\u5c06\u7a00\u758f\u6027\u5f15\u5165\u6df1\u5ea6\u7ef4\u5ea6\uff0c\u4ed6\u4eec\u7684\u505a\u6cd5\u8981\u4e48\u662f\u968f\u673a\u4e22\u5f03\u67d0\u4e9b\u5c42\uff0c\u8981\u4e48\u5c31\u662f\u4f7f\u7528\u53ef\u5b66\u4e60\u7684\u8def\u7531\u5668\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u7684\u505a\u6cd5\u53c2\u8003\u4e86\u7b2c\u4e8c\u79cd\u65b9\u6cd5\uff0c\u540c\u65f6\u6574\u5408\u4e86\u8fd1\u671f\u63d0\u51fa\u7684\u6df7\u5408\u6df1\u5ea6\uff08MoD\uff09\u6280\u672f\u3002\u6709\u5173 MoD \u7684\u66f4\u591a\u4ecb\u7ecd\u53ef\u53c2\u9605\u673a\u5668\u4e4b\u5fc3\u62a5\u9053<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650914627&amp;idx=4&amp;sn=3578ad79a9fe9ab31be7ffe594b8a91d&amp;chksm=84e47b3db393f22b9124b62621b54fad3a02691581f4ef5d0d6e95068cf0050984d2cdf117ad&amp;scene=21#wechat_redirect\" target=\"_blank\">\u300aDeepMind \u5347\u7ea7 Transformer\uff0c\u524d\u5411\u901a\u8fc7 FLOPs \u6700\u591a\u53ef\u964d\u4e00\u534a\u300b<\/a>\u3002<\/section>\n<section><\/section>\n<section>\u5177\u4f53\u800c\u8a00\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8be5\u56e2\u961f\u7684\u505a\u6cd5\u662f\u5728\u6bcf\u4e2a MoD \u5c42\u4e2d\uff0c\u5728\u6df7\u5408\u4e13\u5bb6 \uff08MoE\uff09\u8def\u7531\u4e4b\u524d\u90fd\u96c6\u6210 MoD\uff0c\u4ece\u800c\u786e\u4fdd\u5728\u6a21\u6001\u5206\u79bb\u4e4b\u524d\uff0c\u6574\u6279\u6570\u636e\u90fd\u80fd\u5e94\u7528 MoD\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-7fd24df34ee8c345b7bef1fb89729ed4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-7fd24df34ee8c345b7bef1fb89729ed4.png\"><\/a><\/p>\n<\/section>\n<section><strong>\u63a8\u7406<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u6211\u4eec\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528 MoE \u7684\u4e13\u5bb6\u9009\u62e9\u8def\u7531\u6216 MoD \u7684\u5c42\u9009\u62e9\u8def\u7531\uff0c\u56e0\u4e3a\u5728\u4e00\u6279\u6570\u636e\u4e2d\u8fdb\u884c top-k\uff08\u9009\u62e9\u524d k \u4e2a\uff09\u9009\u62e9\u4f1a\u7834\u574f\u56e0\u679c\u5173\u7cfb\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u4fdd\u8bc1\u63a8\u7406\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u53d7\u4e0a\u8ff0 MoD \u8bba\u6587\u7684\u542f\u53d1\uff0c\u7814\u7a76\u56e2\u961f\u5f15\u5165\u4e86\u8f85\u52a9\u8def\u7531\u5668\uff08auxiliary router\uff09\uff0c\u5176\u4f5c\u7528\u662f\u4ec5\u57fa\u4e8e token \u7684\u9690\u85cf\u8868\u5f81\u9884\u6d4b\u8be5 token \u88ab\u67d0\u4e2a\u4e13\u5bb6\u6216\u5c42\u9009\u4e2d\u7684\u53ef\u80fd\u6027\u3002<\/section>\n<section><\/section>\n<section><strong>\u5347\u7ea7\u6539\u9020\uff08Upcycling\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u4f18\u5316\u8868\u5f81\u7a7a\u95f4\u548c\u8def\u7531\u673a\u5236\u65b9\u9762\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3 MoE \u67b6\u6784\uff0c\u5b58\u5728\u4e00\u4e2a\u72ec\u7279\u7684\u96be\u9898\u3002\u8be5\u56e2\u961f\u53d1\u73b0\uff1aMoE \u8def\u7531\u5668\u8d1f\u8d23\u4e3a\u6bcf\u4e2a\u4e13\u5bb6\u5212\u5206\u8868\u5f81\u7a7a\u95f4\u3002\u4f46\u662f\uff0c\u5728\u6a21\u578b\u8bad\u7ec3\u7684\u65e9\u671f\u9636\u6bb5\uff0c\u8fd9\u4e2a\u8868\u5f81\u7a7a\u95f4\u5e76\u975e\u6700\u4f18\uff0c\u8fd9\u5c31\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u5f97\u5230\u7684\u8def\u7531\u51fd\u6570\u4e5f\u662f\u6b21\u4f18\u7684\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u514b\u670d\u8fd9\u4e00\u5c40\u9650\uff0c\u4ed6\u4eec\u57fa\u4e8e Komatsuzaki \u7b49\u4eba\u7684\u8bba\u6587\u300aSparse upcycling: Training mixture-of-experts from dense checkpoints\u300b\u63d0\u51fa\u4e86\u4e00\u79cd\u5347\u7ea7\u6539\u9020\u65b9\u6cd5\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-38a1ee91574378b2c8c126de38988102.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-38a1ee91574378b2c8c126de38988102.png\"><\/a><\/p>\n<section>\u5177\u4f53\u6765\u8bf4\uff0c\u9996\u5148\u8bad\u7ec3\u4e00\u4e2a\u6bcf\u4e2a\u6a21\u6001\u90fd\u6709\u4e00\u4e2a FFN \u4e13\u5bb6\u7684\u67b6\u6784\u3002\u7ecf\u8fc7\u4e00\u4e9b\u9884\u5148\u8bbe\u5b9a\u7684\u6b65\u6570\u4e4b\u540e\uff0c\u518d\u5bf9\u8be5\u6a21\u578b\u8fdb\u884c\u5347\u7ea7\u6539\u9020\uff0c\u5177\u4f53\u505a\u6cd5\u662f\uff1a\u5c06\u6bcf\u4e2a\u7279\u5b9a\u6a21\u6001\u7684 FFN \u8f6c\u6362\u6210\u4e00\u4e2a\u4e13\u5bb6\u9009\u62e9\u5f0f MoE \u6a21\u5757\uff0c\u5e76\u5c06\u6bcf\u4e2a\u4e13\u5bb6\u521d\u59cb\u5316\u4e3a\u7b2c\u4e00\u9636\u6bb5\u8bad\u7ec3\u7684\u4e13\u5bb6\u3002\u8fd9\u91cc\u4f1a\u5728\u4fdd\u7559\u524d\u4e00\u9636\u6bb5\u7684\u6570\u636e\u52a0\u8f7d\u5668\u72b6\u6001\u7684\u540c\u65f6\u91cd\u7f6e\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\uff0c\u4ee5\u786e\u4fdd\u7b2c\u4e8c\u9636\u6bb5\u7684\u8bad\u7ec3\u80fd\u4f7f\u7528\u5df2\u5237\u65b0\u7684\u6570\u636e\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u4fc3\u8fdb\u4e13\u5bb6\u66f4\u52a0\u4e13\u4e1a\uff0c\u8be5\u56e2\u961f\u8fd8\u4f7f\u7528\u4e86 Gumbel \u566a\u58f0\u6765\u589e\u5f3a MoE \u8def\u7531\u51fd\u6570\uff0c\u4ece\u800c\u4f7f\u5f97\u65b0\u7684\u8def\u7531\u5668\u80fd\u4ee5\u53ef\u5fae\u5206\u7684\u65b9\u5f0f\u5bf9\u4e13\u5bb6\u8fdb\u884c\u91c7\u6837\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u79cd\u5347\u7ea7\u6539\u9020\u65b9\u6cd5\u52a0\u4e0a Gumbel-Sigmoid \u6280\u672f\uff0c\u53ef\u514b\u670d\u5b66\u4e60\u5230\u7684\u8def\u7531\u5668\u7684\u5c40\u9650\u6027\uff0c\u4ece\u800c\u63d0\u5347\u65b0\u63d0\u51fa\u7684\u6a21\u6001\u611f\u77e5\u578b\u7a00\u758f\u67b6\u6784\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>\u6548\u7387\u4f18\u5316<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4fc3\u8fdb MoMa \u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u8be5\u56e2\u961f\u91c7\u7528\u4e86\u5b8c\u5168\u5206\u7247\u5f0f\u6570\u636e\u5e76\u884c\uff08FSDP\/Fully Sharded Data Parallel\uff09\u3002\u4f46\u662f\uff0c\u76f8\u6bd4\u4e8e\u5e38\u89c4 MoE\uff0c\u8be5\u65b9\u6cd5\u5b58\u5728\u4e00\u4e9b\u7279\u6709\u7684\u6548\u7387\u96be\u9898\uff0c\u5305\u62ec\u8d1f\u8f7d\u5e73\u8861\u95ee\u9898\u548c\u4e13\u5bb6\u6267\u884c\u7684\u6548\u7387\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u8d1f\u8f7d\u5e73\u8861\u95ee\u9898\uff0c\u8be5\u56e2\u961f\u5f00\u53d1\u4e86\u4e00\u79cd\u5e73\u8861\u7684\u6570\u636e\u6df7\u5408\u65b9\u6cd5\uff0c\u53ef\u8ba9\u6bcf\u53f0 GPU \u4e0a\u7684\u6587\u672c &#8211; \u56fe\u50cf\u6570\u636e\u6bd4\u4f8b\u4e0e\u4e13\u5bb6\u6bd4\u4f8b\u4fdd\u6301\u4e00\u81f4\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u4e13\u5bb6\u6267\u884c\u7684\u6548\u7387\u95ee\u9898\uff0c\u8be5\u56e2\u961f\u63a2\u7d22\u4e86\u4e00\u4e9b\u7b56\u7565\uff0c\u53ef\u5e2e\u52a9\u63d0\u5347\u4e0d\u540c\u6a21\u6001\u7684\u4e13\u5bb6\u7684\u6267\u884c\u6548\u7387\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5c06\u5404\u4e2a\u6a21\u6001\u7684\u4e13\u5bb6\u9650\u5236\u4e3a\u540c\u6784\u7684\u4e13\u5bb6\uff0c\u5e76\u7981\u6b62\u5c06\u6587\u672c token \u8def\u7531\u5230\u56fe\u50cf\u4e13\u5bb6\uff0c\u53cd\u4e4b\u4ea6\u7136\uff1b<\/section>\n<\/li>\n<li>\n<section>\u4f7f\u7528\u6a21\u5757\u7a00\u758f\u6027\uff08block sparsity\uff09\u6765\u63d0\u5347\u6267\u884c\u6548\u7387\uff1b<\/section>\n<\/li>\n<li>\n<section>\u5f53\u6a21\u6001\u7684\u6570\u91cf\u6709\u9650\u65f6\uff0c\u6309\u987a\u5e8f\u8fd0\u884c\u4e0d\u540c\u6a21\u6001\u7684\u4e13\u5bb6\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u7531\u4e8e\u5b9e\u9a8c\u4e2d\u6bcf\u53f0 GPU \u5904\u7406\u7684 token \u90fd\u8db3\u591f\u591a\uff0c\u56e0\u6b64\u5373\u4f7f\u4f7f\u7528\u591a\u4e2a\u5206\u6279\u6b21\u77e9\u9635\u4e58\u6cd5\uff0c\u786c\u4ef6\u5229\u7528\u7387\u4e5f\u4e0d\u7b97\u5927\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u8be5\u56e2\u961f\u8ba4\u4e3a\u5bf9\u4e8e\u5f53\u524d\u89c4\u6a21\u7684\u5b9e\u9a8c\u73af\u5883\u800c\u8a00\uff0c\u6309\u987a\u5e8f\u6267\u884c\u7684\u65b9\u6cd5\u662f\u6bd4\u8f83\u597d\u7684\u9009\u62e9\u3002<\/section>\n<section><\/section>\n<section><strong>\u5176\u5b83\u4f18\u5316<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\u541e\u5410\u91cf\uff0c\u8be5\u56e2\u961f\u8fd8\u91c7\u7528\u4e86\u5176\u5b83\u4e00\u4e9b\u4f18\u5316\u6280\u672f\u3002<\/section>\n<section><\/section>\n<section>\u5176\u4e2d\u5305\u62ec\u964d\u4f4e\u68af\u5ea6\u901a\u4fe1\u91cf\u3001\u81ea\u52a8\u5316\u7684 GPU \u6838\u878d\u5408\u7b49\u4e00\u822c\u4f18\u5316\u64cd\u4f5c\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u901a\u8fc7 torch.compile \u5b9e\u73b0\u4e86\u56fe\u4f18\u5316\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u9488\u5bf9 MoMa \u5f00\u53d1\u4e86\u4e00\u4e9b\u4f18\u5316\u6280\u672f\uff0c\u5305\u62ec\u8de8\u4e0d\u540c\u5c42\u590d\u7528\u6a21\u6001 token \u7d22\u5f15\uff0c\u4ee5\u6700\u9ad8\u6548\u5730\u540c\u6b65 CPU \u548c GPU \u4e4b\u95f4\u7684\u8bbe\u5907\u3002<\/section>\n<section><\/section>\n<section><strong>\u5b9e\u9a8c<\/strong><\/section>\n<section><\/section>\n<section><strong>\u8bbe\u7f6e<\/strong><\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u548c\u9884\u5904\u7406\u8fc7\u7a0b\u4e0e Chameleon \u4e00\u6837\u3002\u4e3a\u4e86\u8bc4\u4f30\u6269\u5c55\u6027\u80fd\uff0c\u4ed6\u4eec\u8bad\u7ec3\u6a21\u578b\u4f7f\u7528\u7684 token \u6570\u91cf\u8d85\u8fc7 1 \u4e07\u4ebf\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4ddba255b4b028b5b23e51baae687975.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4ddba255b4b028b5b23e51baae687975.png\"><\/a><\/p>\n<\/section>\n<section>\u8868 1 \u7ed9\u51fa\u4e86\u5bc6\u96c6\u548c\u7a00\u758f\u6a21\u578b\u7684\u8be6\u7ec6\u914d\u7f6e\u60c5\u51b5\u3002<\/section>\n<section><\/section>\n<section><strong>\u4e0d\u540c\u8ba1\u7b97\u5c42\u7ea7\u7684\u6269\u5c55\u6027\u80fd<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u5206\u6790\u4e86\u4e0d\u540c\u6a21\u578b\u5728\u4e0d\u540c\u8ba1\u7b97\u5c42\u7ea7\u4e0a\u7684\u6269\u5c55\u6027\u80fd\uff0c\u8fd9\u4e9b\u8ba1\u7b97\u5c42\u7ea7\uff08FLOPs\uff09\u76f8\u5f53\u4e8e\u4e09\u79cd\u5927\u5c0f\u7684\u5bc6\u96c6\u6a21\u578b\uff1a90M\u3001435M \u548c 1.4B\u3002<\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u4e00\u4e2a\u7a00\u758f\u6a21\u578b\u4ec5\u4f7f\u7528\u603b FLOPs \u7684 1\/\u03b7 \u5c31\u80fd\u6bd4\u80a9\u540c\u7b49 FLOPs \u7684\u5bc6\u96c6\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u635f\u5931\uff08\u03b7 \u8868\u793a\u9884\u8bad\u7ec3\u52a0\u901f\u56e0\u5b50\uff09\u3002<\/section>\n<section><\/section>\n<section><strong>\u6a21\u6001\u89e3\u7ed1<\/strong><\/section>\n<section><\/section>\n<section>\u5f15\u5165\u7279\u5b9a\u6a21\u6001\u7684\u4e13\u5bb6\u5206\u7ec4\u53ef\u63d0\u9ad8\u4e0d\u540c\u89c4\u6a21\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6548\u7387\uff0c\u8fd9\u5bf9\u56fe\u50cf\u6a21\u6001\u5c24\u5176\u6709\u76ca\u3002\u5982\u56fe 3 \u6240\u793a\uff0c\u4f7f\u7528 1 \u4e2a\u56fe\u50cf\u4e13\u5bb6\u548c 1 \u4e2a\u6587\u672c\u4e13\u5bb6\u7684 moe_1t1i \u914d\u7f6e\u663e\u8457\u4f18\u4e8e\u76f8\u5e94\u7684\u5bc6\u96c6\u6a21\u578b\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-068179b73a2d3c551d9d23597dce2b0a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-068179b73a2d3c551d9d23597dce2b0a.png\"><\/a><\/p>\n<\/section>\n<section>\u6269\u5c55\u6bcf\u4e2a\u6a21\u6001\u5206\u7ec4\u7684\u4e13\u5bb6\u6570\u91cf\u8fd8\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>\u6df7\u5408\u6df1\u5ea6\u4e0e\u4e13\u5bb6<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u89c2\u5bdf\u5230\uff0c\u5f53\u91c7\u7528 MoE \u548c MoD \u4ee5\u53ca\u5b83\u4eec\u7684\u7ec4\u5408\u5f62\u5f0f\u65f6\uff0c\u8bad\u7ec3\u635f\u5931\u7684\u6536\u655b\u901f\u5ea6\u4f1a\u5f97\u5230\u63d0\u5347\u3002\u5982\u56fe 4 \u6240\u793a\uff0c\u5411 moe_1t1i \u67b6\u6784\u6dfb\u52a0 MoD\uff08mod_moe_1t1i\uff09\u53ef\u5927\u5e45\u63d0\u5347\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u7684\u6a21\u578b\u6027\u80fd\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-812e85198127ac77bb995a4998f7ca57.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-812e85198127ac77bb995a4998f7ca57.png\"><\/a><\/p>\n<\/section>\n<section>\u6b64\u5916\uff0c\u5728\u4e0d\u540c\u7684\u6a21\u578b\u5927\u5c0f\u548c\u6a21\u6001\u4e0a\uff0cmod_moe_1t1i \u80fd\u5ab2\u7f8e\u751a\u81f3\u8d85\u8fc7 moe_4t4i\uff0c\u8fd9\u8868\u660e\u5728\u6df1\u5ea6\u7ef4\u5ea6\u4e0a\u5f15\u5165\u7a00\u758f\u6027\u4e5f\u80fd\u6709\u6548\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u3002<\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u65b9\u9762\uff0c\u8fd8\u80fd\u770b\u5230\u5806\u53e0 MoD \u548c MoE \u7684\u6536\u76ca\u4f1a\u9010\u6b65\u4e0b\u964d\u3002<\/section>\n<section><\/section>\n<section><strong>\u6269\u5c55\u4e13\u5bb6\u7684\u6570\u91cf<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u7814\u7a76\u6269\u5c55\u4e13\u5bb6\u6570\u91cf\u7684\u5f71\u54cd\uff0c\u8be5\u56e2\u961f\u8fdb\u884c\u4e86\u8fdb\u4e00\u6b65\u7684\u6d88\u878d\u5b9e\u9a8c\u3002\u4ed6\u4eec\u63a2\u7d22\u4e86\u4e24\u79cd\u573a\u666f\uff1a\u4e3a\u6bcf\u79cd\u6a21\u6001\u5206\u914d\u540c\u7b49\u6570\u91cf\u7684\u4e13\u5bb6\uff08\u5e73\u8861\uff09\u4ee5\u53ca\u4e3a\u6bcf\u79cd\u6a21\u6001\u5206\u914d\u4e0d\u540c\u6570\u91cf\u7684\u4e13\u5bb6\uff08\u4e0d\u5e73\u8861\uff09\u3002\u7ed3\u679c\u89c1\u56fe 5\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-ac02d425a0b6bed2adae70daeab70cec.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-ac02d425a0b6bed2adae70daeab70cec.png\"><\/a><\/p>\n<\/section>\n<section>\u5bf9\u4e8e\u5e73\u8861\u7684\u8bbe\u7f6e\uff0c\u4ece\u56fe 5a \u53ef\u4ee5\u770b\u5230\uff0c\u968f\u7740\u4e13\u5bb6\u6570\u91cf\u63d0\u5347\uff0c\u8bad\u7ec3\u635f\u5931\u4f1a\u660e\u663e\u4e0b\u964d\u3002\u4f46\u6587\u672c\u548c\u56fe\u50cf\u635f\u5931\u8868\u73b0\u51fa\u4e86\u4e0d\u540c\u7684\u6269\u5c55\u6a21\u5f0f\u3002\u8fd9\u8868\u660e\u6bcf\u79cd\u6a21\u6001\u7684\u56fa\u6709\u7279\u6027\u4f1a\u5bfc\u81f4\u51fa\u73b0\u4e0d\u540c\u7684\u7a00\u758f\u5efa\u6a21\u884c\u4e3a\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u4e0d\u5e73\u8861\u7684\u8bbe\u7f6e\uff0c\u56fe 5b \u6bd4\u8f83\u4e86\u540c\u7b49\u4e13\u5bb6\u603b\u6570\uff088\uff09\u7684\u4e09\u79cd\u4e0d\u540c\u914d\u7f6e\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u4e00\u4e2a\u6a21\u6001\u7684\u4e13\u5bb6\u8d8a\u591a\uff0c\u6a21\u578b\u5728\u8be5\u6a21\u6001\u4e0a\u7684\u8868\u73b0\u901a\u5e38\u5c31\u8d8a\u597d\u3002<\/section>\n<section><\/section>\n<section><strong>\u5347\u7ea7\u6539\u9020<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u81ea\u7136\u4e5f\u9a8c\u8bc1\u4e86\u524d\u8ff0\u7684\u5347\u7ea7\u6539\u9020\u7684\u6548\u679c\u3002\u56fe 6 \u6bd4\u8f83\u4e86\u4e0d\u540c\u6a21\u578b\u53d8\u4f53\u7684\u8bad\u7ec3\u66f2\u7ebf\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-09a540675f5b15efa904eadcbabd8d02.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-09a540675f5b15efa904eadcbabd8d02.png\"><\/a><\/p>\n<\/section>\n<section>\u7ed3\u679c\u8868\u660e\uff0c\u5347\u7ea7\u6539\u9020\u786e\u5b9e\u80fd\u8fdb\u4e00\u6b65\u6539\u5584\u6a21\u578b\u8bad\u7ec3\uff1a\u5f53\u7b2c\u4e00\u4e2a\u9636\u6bb5\u6709 10k \u6b65\u65f6\uff0c\u5347\u7ea7\u6539\u9020\u80fd\u5e26\u6765 1.2 \u500d\u7684 FLOPs \u6536\u76ca\uff1b\u800c\u5f53\u8fd9\u4e2a\u6b65\u6570\u4e3a 20k \u65f6\uff0c\u4e5f\u6709 1.16 \u500d\u7684 FLOPs \u6536\u76ca\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u8fd8\u80fd\u89c2\u5bdf\u5230\uff0c\u968f\u7740\u8bad\u7ec3\u63a8\u8fdb\uff0c\u7ecf\u8fc7\u5347\u7ea7\u6539\u9020\u7684\u6a21\u578b\u4e0e\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u7684\u6a21\u578b\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u4f1a\u4e0d\u65ad\u589e\u5927\u3002<\/section>\n<section><\/section>\n<section><strong>\u541e\u5410\u91cf\u5206\u6790<\/strong><\/section>\n<section><\/section>\n<section>\u7a00\u758f\u6a21\u578b\u901a\u5e38\u4e0d\u80fd\u7acb\u5373\u5e26\u6765\u6027\u80fd\u589e\u76ca\uff0c\u56e0\u4e3a\u7a00\u758f\u6a21\u578b\u4f1a\u589e\u52a0\u52a8\u6001\u6027\u548c\u76f8\u5173\u7684\u6570\u636e\u5e73\u8861\u95ee\u9898\u3002\u4e3a\u4e86\u91cf\u5316\u65b0\u63d0\u51fa\u7684\u65b9\u6cd5\u5bf9\u8bad\u7ec3\u6548\u7387\u7684\u5f71\u54cd\uff0c\u8be5\u56e2\u961f\u901a\u5e38\u63a7\u5236\u53d8\u91cf\u5b9e\u9a8c\u6bd4\u8f83\u4e86\u4e0d\u540c\u67b6\u6784\u7684\u8bad\u7ec3\u541e\u5410\u91cf\u3002\u7ed3\u679c\u89c1\u8868 2\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-114481a8010842ee78addbb00f4e10d1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-114481a8010842ee78addbb00f4e10d1.png\"><\/a><\/p>\n<\/section>\n<section>\u53ef\u4ee5\u770b\u5230\uff0c\u76f8\u6bd4\u4e8e\u5bc6\u96c6\u6a21\u578b\uff0c\u57fa\u4e8e\u6a21\u6001\u7684\u7a00\u758f\u6027\u80fd\u5b9e\u73b0\u66f4\u597d\u7684\u8d28\u91cf &#8211; \u541e\u5410\u91cf\u6743\u8861\uff0c\u5e76\u4e14\u80fd\u968f\u4e13\u5bb6\u6570\u91cf\u589e\u957f\u5c55\u73b0\u51fa\u5408\u7406\u7684\u53ef\u6269\u5c55\u6027\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5c3d\u7ba1 MoD \u53d8\u4f53\u53d6\u5f97\u4e86\u6700\u597d\u7684\u7edd\u5bf9\u635f\u5931\uff0c\u4f46\u7531\u4e8e\u989d\u5916\u7684\u52a8\u6001\u6027\u548c\u4e0d\u5e73\u8861\u6027\uff0c\u5b83\u4eec\u7684\u8ba1\u7b97\u6210\u672c\u5f80\u5f80\u4e5f\u66f4\u9ad8\u3002<\/section>\n<section><\/section>\n<section><strong>\u63a8\u7406\u65f6\u95f4\u6027\u80fd<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u4e5f\u8bc4\u4f30\u4e86\u6a21\u578b\u5728\u7559\u5b58\u7684\u8bed\u8a00\u5efa\u6a21\u6570\u636e\u548c\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002\u7ed3\u679c\u89c1\u8868 3 \u548c 4\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e938dcde886dec7808c59f420a781642.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e938dcde886dec7808c59f420a781642.png\"><\/a><\/section>\n<section>\u5982\u8868 3 \u6240\u793a\uff0c\u901a\u8fc7\u4f7f\u7528\u591a\u4e2a\u56fe\u50cf\u4e13\u5bb6\uff0c1.4B MoMa 1t1i \u6a21\u578b\u5728\u5927\u591a\u6570\u6307\u6807\u4e0a\u90fd\u4f18\u4e8e\u76f8\u5e94\u7684\u5bc6\u96c6\u6a21\u578b\uff0c\u53ea\u6709\u5728 COCO \u548c Flickr \u4e0a\u7684\u56fe\u50cf\u5230\u6587\u672c\u6761\u4ef6\u56f0\u60d1\u5ea6\u6307\u6807\u4f8b\u5916\u3002\u8fdb\u4e00\u6b65\u6269\u5c55\u4e13\u5bb6\u6570\u91cf\u4e5f\u80fd\u63d0\u5347\u6027\u80fd\uff0c\u5176\u4e2d 1.4B MoE 8x \u5728\u56fe\u50cf\u5230\u6587\u672c\u6027\u80fd\u4e0a\u8fbe\u5230\u4e86\u6700\u4f73\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u5982\u8868 4 \u6240\u793a\uff0c1.4B MoE 8x \u8fd9\u4e2a\u6a21\u578b\u8fd8\u975e\u5e38\u64c5\u957f\u6587\u672c\u5230\u6587\u672c\u4efb\u52a1\u30021.4B MoMa 4t4i \u5728\u6240\u6709\u6761\u4ef6\u56fe\u50cf\u56f0\u60d1\u5ea6\u6307\u6807\u4e0a\u8868\u73b0\u6700\u4f73\uff0c\u800c\u5176\u5728\u5927\u591a\u6570\u57fa\u51c6\u4e0a\u7684\u6587\u672c\u56f0\u60d1\u5ea6\u4e5f\u975e\u5e38\u63a5\u8fd1 1.4B MoE 8x\u3002<\/section>\n<section><\/section>\n<section>\u603b\u4f53\u800c\u8a00\uff0c\u5728\u6df7\u5408\u6587\u672c\u548c\u56fe\u50cf\u4e24\u79cd\u6a21\u6001\u7684\u6570\u636e\u4e0a\uff0c1.4B MoMa 4t4i \u6a21\u578b\u7684\u5efa\u6a21\u7ed3\u679c\u6700\u597d\u3002<\/section>\n<section><\/section>\n<section>\u66f4\u591a\u8be6\u7ec6\u5185\u5bb9\uff0c\u8bf7\u9605\u8bfb\u539f\u8bba\u6587\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-11-2\" target=\"_blank\">\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd [&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":[73],"class_list":["post-35037","post","type-post","status-publish","format-standard","hentry","category-news","tag-73"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408 - \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=35037\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=35037\",\"url\":\"https:\/\/17aitech.com\/?p=35037\",\"name\":\"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=35037#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=35037#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg\",\"datePublished\":\"2024-12-17T02:59:37+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=35037#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=35037\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=35037#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=35037#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408\"}]},{\"@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\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"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":"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408 - \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=35037","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35037","url":"https:\/\/17aitech.com\/?p=35037","name":"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35037#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35037#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg","datePublished":"2024-12-17T02:59:37+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35037#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35037"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35037#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-58cd3280426d4184a6442c1cb8b95480.jpeg"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35037#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u6df7\u5408\u4e13\u5bb6\u66f4\u6709\u4e3b\u89c1\u4e86\uff0c\u80fd\u611f\u77e5\u591a\u6a21\u6001\u5206\u60c5\u51b5\u884c\u4e8b\uff0cMeta\u63d0\u51fa\u6a21\u6001\u611f\u77e5\u578b\u4e13\u5bb6\u6df7\u5408"}]},{"@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":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","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\/35037","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=35037"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35037\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35037"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35037"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35037"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}