{"id":24872,"date":"2025-03-05T10:08:44","date_gmt":"2025-03-05T02:08:44","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=24872"},"modified":"2025-03-05T10:08:44","modified_gmt":"2025-03-05T02:08:44","slug":"%e4%b8%badeepseek-moe%e6%a8%a1%e5%9e%8b%e5%b8%a6%e6%9d%a5%e3%80%8c%e5%85%8d%e8%b4%b9%e5%8d%88%e9%a4%90%e3%80%8d%e5%8a%a0%e9%80%9f%ef%bc%8c%e4%b8%93%e5%ae%b6%e9%93%be%e5%8f%af%e5%a4%a7%e5%b9%85","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/03\/05\/24872\/","title":{"rendered":"\u4e3aDeepSeek MoE\u6a21\u578b\u5e26\u6765\u300c\u514d\u8d39\u5348\u9910\u300d\u52a0\u901f\uff0c\u4e13\u5bb6\u94fe\u53ef\u5927\u5e45\u63d0\u5347LLM\u7684\u4fe1\u606f\u5904\u7406\u80fd\u529b"},"content":{"rendered":"
\"\"<\/div>\n
<\/div>\n
Chain-of-Experts \u4ee3\u8868\u4e86\u9ad8\u6548\u3001\u9ad8\u6027\u80fd\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u7684\u91cd\u8981\u4e00\u6b65\u3002<\/div>\n
\n
\n

\u6211\u4eec\u90fd\u77e5\u9053\uff0cDeepSeek-R1 \u7684\u8bad\u7ec3\u8fc7\u7a0b\u4f7f\u7528\u4e86\u4e00\u79cd\u540d\u4e3a\u4e13\u5bb6\u6df7\u5408\u6a21\u578b\uff08Mixture-of-Experts, MoE\uff09\u7684\u6280\u672f\uff0c\u800c\u5f53\u524d\u7684 MoE \u6280\u672f\u4f9d\u7136\u8fd8\u6709\u663e\u8457\u7684\u4f18\u5316\u7a7a\u95f4\u3002<\/p>\n

\u8fd1\u65e5\uff0c\u7f8e\u56fd\u897f\u5317\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u535a\u58eb\u751f\u738b\u5b50\u6db5\uff08Zihan Wang\uff09\u7b49\u4eba\u5728\u8fd9\u4e2a\u65b9\u5411\u4e0a\u53d6\u5f97\u4e86\u7a81\u7834\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u4e13\u5bb6\u94fe\uff08CoE\uff09<\/strong>\u7684\u6280\u672f\u3002\u5b9e\u9a8c\u8868\u660e\uff0cCoE \u5728\u6027\u80fd\u3001\u6269\u5c55\u7b56\u7565\u3001\u8d44\u6e90\u6548\u7387\u548c\u4e13\u5bb6\u4f7f\u7528\u6548\u7387\u7b49\u591a\u4e2a\u65b9\u9762\u90fd\u663e\u8457\u8d85\u8d8a\u5148\u524d\u7684 MoE \u6a21\u578b\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u76ee\u524d\uff0c\u6838\u5fc3\u7814\u7a76\u8005\u738b\u5b50\u6db5\u5df2\u7ecf\u64b0\u5199\u5e76\u53d1\u5e03 CoE \u7684\u4e2d\u82f1\u6587\u535a\u5ba2\u5e76\u5df2\u7ecf\u5728 GitHub \u5f00\u6e90\u76f8\u5173\u4ee3\u7801\uff0c\u4ed6\u4e5f\u5728\u535a\u5ba2\u4e2d\u8868\u793a\u7814\u7a76\u8bba\u6587\u5c06\u5728\u8fd1\u671f\u53d1\u5e03\u3002\u4ee5\u4e0b\u4e3a\u76f8\u5173\u94fe\u63a5\uff1a<\/p>\n

    \n
  • \u4ee3\u7801\uff1ahttps:\/\/github.com\/ZihanWang314\/coe<\/li>\n
  • \u4e2d\u6587\u62a5\u544a\uff1ahttps:\/\/sandy-server-87f.notion.site\/1ab9bb750b79801bbfebf01ae9a77b3f<\/li>\n
  • \u82f1\u6587\u62a5\u544a\uff1ahttps:\/\/sandy-server-87f.notion.site\/Chain-of-Experts-Unlocking-the-Communication-Power-of-MoEs-1ab9bb750b7980048d43e6aab3537cea<\/li>\n<\/ul>\n

    \u673a\u5668\u4e4b\u5fc3\u6388\u6743\u8f6c\u8f7d\u4e86\u5176\u4e2d\u6587\u62a5\u544a\u3002<\/p>\n

    \u4f5c\u8005\u7b80\u4ecb\uff1a\u738b\u5b50\u6db5<\/strong>\uff0c\u7f8e\u56fd\u897f\u5317\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u535a\u58eb\u4e00\u5e74\u7ea7\u5b66\u751f\uff0c\u5bfc\u5e08\u4e3a Manling Li\u3002\u4ed6\u672c\u79d1\u6bd5\u4e1a\u4e8e\u4e2d\u56fd\u4eba\u6c11\u5927\u5b66\u9ad8\u74f4\u4eba\u5de5\u667a\u80fd\u5b66\u9662\uff0c\u5e76\u66fe\u53c2\u4e0e\u8fc7 DeepSeek-V2 \u6a21\u578b\u7684\u5f00\u53d1\u5de5\u4f5c\u3002<\/p><\/blockquote>\n

    \u5f15\u8a00<\/h3>\n

    \u6211\u4eec\u63d0\u51fa\u4e13\u5bb6\u94fe (Chain-of-Experts\uff0cCoE) \u67b6\u6784\uff0c\u4e00\u79cd\u901a\u8fc7\u5728\u5355\u5c42\u5185\u5b9e\u73b0\u4e13\u5bb6\u95f4\u4e32\u884c\u901a\u4fe1\u7684\u521b\u65b0\u65b9\u6cd5\uff0c\u4ece\u6839\u672c\u4e0a\u6539\u53d8\u7a00\u758f\u795e\u7ecf\u7f51\u7edc\u7684\u4fe1\u606f\u5904\u7406\u65b9\u5f0f\u3002<\/p>\n

    MoE \u8bbe\u8ba1\u4e2d\u5b58\u5728\u4e13\u5bb6\u95f4\u72ec\u7acb\u5904\u7406\u4ee5\u53ca\u663e\u5b58\u9700\u6c42\u9ad8\u7684\u95ee\u9898\u3002\u4e0e\u5148\u524d MoE \u72ec\u7acb\u5904\u7406\u6bcf\u4e2a token \u4e0d\u540c\uff0cCoE \u5f15\u5165\u8fed\u4ee3\u673a\u5236\u4f7f\u4e13\u5bb6\u80fd\u591f “\u6c9f\u901a”\uff0c\u5728\u5176\u4ed6\u4e13\u5bb6\u7684\u8f93\u51fa\u4e4b\u4e0a\u5904\u7406 token\u3002<\/p>\n

    \u5b9e\u9a8c\u8bc1\u660e CoE \u5728\u591a\u4e2a\u65b9\u9762\u663e\u8457\u8d85\u8d8a\u5148\u524d MoE \u6a21\u578b\u3002\u6027\u80fd\u663e\u8457\u63d0\u5347<\/strong>\uff0cCoE \u4e24\u6b21\u8fed\u4ee3\u5728 Math \u4efb\u52a1\u4e0a\u5c06\u9a8c\u8bc1 loss \u4ece 1.20 \u964d\u81f3 1.12\uff1b\u6269\u5c55\u7b56\u7565\u4f18\u52bf<\/strong>\uff0cCoE \u4e2d\u6269\u5c55\u4e24\u6b21\u8fed\u4ee3\u6027\u80fd\u76f8\u5f53\u4e8e\u5c06\u4e13\u5bb6\u9009\u62e9\u6570\u91cf\u6269\u5927 3 \u500d\uff0c\u5e76\u4f18\u4e8e\u6269\u5c55\u6a21\u578b\u5c42\u6570\uff1b\u8d44\u6e90\u6548\u7387\u4f18\u5316<\/strong>\uff0c\u76f8\u4f3c\u8868\u73b0\u4e0b\u51cf\u5c11\u4e86 17.6-42% \u7684\u5185\u5b58\u9700\u6c42\uff1b\u4e13\u5bb6\u7ec4\u5408\u81ea\u7531\u5ea6<\/strong>\u63d0\u5347\u4e86 823 \u500d\uff1b\u4e13\u5bb6\u4f7f\u7528\u6548\u7387<\/strong>\u589e\u5f3a\uff0c\u4fc3\u8fdb\u4e86\u4e13\u5bb6\u95f4\u7684\u76f4\u63a5\u901a\u4fe1\u548c\u66f4\u597d\u7684\u4e13\u5bb6\u5206\u5316\u3002\u8fd9\u4e9b\u4f18\u52bf\u6784\u6210\u4e86\u4e00\u79cd\u7f55\u89c1\u7684 “\u514d\u8d39\u5348\u9910” \u6548\u5e94\uff0c\u4e3a\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u9ad8\u6548\u6269\u5c55\u5f00\u8f9f\u4e86\u65b0\u9014\u5f84\u3002<\/p>\n

    Chain-of-Experts\uff1a\u91ca\u653e MoE \u4e13\u5bb6\u7684\u6c9f\u901a\u6f5c\u80fd<\/h3>\n

    \u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u4e0d\u65ad\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u53ef\u80fd\u6027\u7684\u8fb9\u754c\uff0c\u4f46\u9ad8\u6548\u6269\u5c55\u8fd9\u4e9b\u6a21\u578b\u4ecd\u7136\u662f\u4e00\u4e2a\u91cd\u5927\u6311\u6218\u3002\u4e13\u5bb6\u6df7\u5408\uff08Mixture of Experts\uff0cMoE\uff09\u6a21\u578b\u4f5c\u4e3a\u4e00\u79cd\u6709\u524d\u666f\u7684\u65b9\u6cd5\u51fa\u73b0\uff0c\u901a\u8fc7\u6bcf\u4e2a token \u4ec5\u6fc0\u6d3b\u90e8\u5206\u53c2\u6570\u6765\u89e3\u51b3\u8fd9\u4e00\u6311\u6218\uff0c\u7406\u8bba\u4e0a\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u6269\u5c55\u3002\u7136\u800c\uff0cMoE \u6a21\u578b\u5b58\u5728\u4ee5\u4e0b\u5c40\u9650\u6027\uff1a<\/p>\n

      \n
    1. \u72ec\u7acb\u6807\u8bb0\u5904\u7406\uff1aMoE \u6a21\u578b\u901a\u5e38\u5e76\u884c\u72ec\u7acb\u5904\u7406 token\uff0c\u4e13\u5bb6\u4e4b\u95f4\u6ca1\u6709\u6c9f\u901a\u3002<\/li>\n
    2. \u5185\u5b58\u6548\u7387\u4f4e\u4e0b\uff1a\u7531\u4e8e\u5177\u6709\u7a00\u758f\u6fc0\u6d3b\u6a21\u5f0f\uff0cMoE \u6574\u4f53\u53c2\u6570\u6570\u91cf\u8f83\u5927\uff0c\u9700\u8981\u5927\u91cf\u5185\u5b58\u8d44\u6e90\u3002<\/li>\n<\/ol>\n

      Chain-of-Experts (CoE) \u4ecb\u7ecd<\/strong><\/p>\n

      \u6211\u4eec\u7684\u7814\u7a76\u5f15\u5165\u4e86 Chain-of-Experts (CoE)\uff0c\u8fd9\u662f\u4e00\u79cd\u6539\u53d8\u7a00\u758f\u795e\u7ecf\u7f51\u7edc\u5904\u7406\u4fe1\u606f\u65b9\u5f0f\u7684\u65b0\u65b9\u6cd5\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      Chain-of-Experts \u7684\u5f62\u5f0f\u5316\u8868\u8ff0<\/h3>\n

      CoE \u7684\u5173\u952e\u521b\u65b0\u5728\u4e8e\u5efa\u7acb\u6c9f\u901a\u6027\u5904\u7406\u673a\u5236\uff0c\u8d85\u8d8a\u4e86\u5148\u524d MoE \u6a21\u578b\u4e2d\u7684\u72ec\u7acb\u6807\u51c6\u5904\u7406\u3002\u6211\u4eec\u5982\u4f55\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff1f\u901a\u8fc7\u5728\u5355\u4e2a\u5c42\u7684\u8fed\u4ee3\u4e2d\u5c06 MoE \u8f93\u51fa\u53cd\u9988\u4e3a\u591a\u6b21\u8fed\u4ee3\u7684\u8f93\u5165\u3002<\/p>\n

      1\u3001\u5148\u524d MoE \u7684\u8f93\u51fa\u8868\u793a<\/strong><\/p>\n

      \u5728\u5148\u524d\u7684 MoE \u5c42\u4e2d\uff0c\u4e0d\u8003\u8651\u5171\u4eab\u4e13\u5bb6\uff0c\u8f93\u51fa\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u5176\u4e2d x \u662f\u8f93\u5165\uff0c\"\u56fe\u7247\"\u00a0\u662f\u95e8\u63a7\u503c\uff08\u8def\u7531\u51fd\u6570\u5206\u914d\u7ed9\u4e13\u5bb6\"\u56fe\u7247\"\u00a0\u7684\u6743\u91cd\uff09\uff0cK \u662f\u6bcf\u4e2a token \u9009\u62e9\u7684\u4e13\u5bb6\u6570\u91cf\uff0cN \u662f\u6bcf\u5c42\u4e13\u5bb6\u603b\u6570\uff0c\"\u56fe\u7247\"\u00a0\u662f\u4e13\u5bb6 i \u7684\u8def\u7531\u53c2\u6570\u3002<\/p>\n

      \u8003\u8651\u5171\u4eab\u4e13\u5bb6\u65f6\uff0c\u8def\u7531\u51fd\u6570\u8ba1\u7b97\u4e0d\u53d8\uff0cMoE \u7684\u8f93\u51fa\u53d8\u4e3a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u5176\u4e2d M \u662f\u5171\u4eab\u4e13\u5bb6\u7684\u603b\u6570\uff0c\"\u56fe\u7247\"\u662f\u5171\u4eab\u4e13\u5bb6\u3002<\/p>\n

      2\u3001CoE \u7684\u8fed\u4ee3\u5904\u7406\u673a\u5236<\/strong><\/p>\n

      \u5728 CoE \u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u8fed\u4ee3\u5904\u7406\u673a\u5236\uff0c\u53ef\u4ee5\u5f62\u5f0f\u5316\u8868\u793a\u4e3a\uff1a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u5176\u4e2d\u00a0\"\u56fe\u7247\"\u00a0\u00a0\u662f\u7b2c t\u00a0\u6b21\u8fed\u4ee3\u7684\u4e2d\u95f4\u8868\u793a\uff0c\"\u56fe\u7247\"\u00a0\u662f\u7b2c t\u00a0\u6b21\u8fed\u4ee3\u4e2d\u4e13\u5bb6 i\u00a0\u7684\u95e8\u63a7\u503c\uff0cC\u00a0\u662f\u8fed\u4ee3\u7684\u603b\u6b21\u6570\uff0c\"\u56fe\u7247\"\u662f\u7b2c i\u00a0\u4e2a\u4e13\u5bb6\uff0c\"\u56fe\u7247\"\u00a0\u662f\u6b8b\u5dee\u8fde\u63a5\u6307\u793a\u53d8\u91cf\uff0c\u5f53\u4f7f\u7528\u5185\u6b8b\u5dee\u65f6\"\u56fe\u7247\"\uff0c\u5426\u5219\u4e3a0\u3002<\/p>\n

      \u8003\u8651\u5230\u5171\u4eab\u4e13\u5bb6\uff0c\u53ef\u4ee5\u5f62\u5f0f\u5316\u8868\u793a\u4e3a\uff1a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      3\u3001\u95e8\u63a7\u673a\u5236\u8be6\u7ec6\u8bf4\u660e<\/strong><\/p>\n

      \u53c2\u8003 DeepSeek-V2 \u7684\u5b9e\u73b0\uff0c\u6211\u4eec\u5b9a\u4e49\u95e8\u63a7\u673a\u5236\u4e3a\uff1a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u5176\u4e2d\uff0c\u6211\u4eec\u4e3a\u6bcf\u6b21\u8fed\u4ee3\u4f7f\u7528\u72ec\u7acb\u95e8\u63a7\uff0c\u91c7\u7528\u4e0d\u540c\u7684\u8def\u7531\u51fd\u6570\uff0cK\/C \u662f\u6bcf\u6b21\u8fed\u4ee3\u9009\u62e9\u7684\u4e13\u5bb6\u6570\u91cf\uff08\u4ee5\u4fdd\u6301\u603b\u8ba1\u7b97\u91cf\u4e0d\u53d8\uff09\uff0c\"\u56fe\u7247\"\u00a0\u662f\u7b2c t \u6b21\u8fed\u4ee3\u4e2d\u4e13\u5bb6 i\u00a0\u7684\u8def\u7531\u53c2\u6570\u3002<\/p>\n

      \u65b9\u6cd5\u5206\u6790<\/h4>\n

      1\u3001CoE \u7684\u4f18\u52bf<\/strong><\/p>\n

      \u8fd9\u79cd\u8fed\u4ee3\u5f0f\u4e13\u5bb6\u5904\u7406\u65b9\u6cd5\u5177\u6709\u4e24\u4e2a\u5173\u952e\u4f18\u52bf\uff1a<\/p>\n

        \n
      1. \u6bcf\u6b21\u8fed\u4ee3\u7684\u4e13\u5bb6\u9009\u62e9\u7531\u524d\u4e00\u6b21\u8fed\u4ee3\u7684\u8f93\u51fa\u51b3\u5b9a\uff0c\u5f62\u6210\u4e13\u5bb6\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u548c\u66f4\u52a8\u6001\u7684\u8def\u7531\u673a\u5236<\/li>\n
      2. \u4e32\u884c\u4fe1\u606f\u53ef\u4ee5\u5728\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\u7d2f\u79ef\uff0c\u5b9e\u73b0\u4e13\u5bb6\u95f4\u7684\u76f4\u63a5\u901a\u4fe1<\/li>\n<\/ol>\n

        \u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0cCoE \u80fd\u591f\u5728\u4fdd\u6301\u8ba1\u7b97\u6548\u7387\u7684\u540c\u65f6\uff0c\u663e\u8457\u63d0\u9ad8\u6a21\u578b\u7684\u8868\u73b0\uff0c\u7279\u522b\u662f\u5728\u590d\u6742\u7684\u573a\u666f\u4e2d\uff08\u5982\u5b9e\u9a8c\u91c7\u7528\u7684\u6570\u5b66\u4efb\u52a1\uff09\u3002<\/p>\n

        2.\u8def\u7531\u51fd\u6570\u5982\u4f55\u5f71\u54cd\u51b3\u7b56\u8fc7\u7a0b<\/strong><\/p>\n

        \u5728 CoE \u6a21\u578b\u4e2d\uff0c\u8def\u7531\u51fd\u6570\"\u56fe\u7247\"\u5bf9\u51b3\u7b56\u8fc7\u7a0b\u7684\u5f71\u54cd\u662f\u591a\u5c42\u6b21\u7684\uff1a<\/p>\n

          \n
        1. \u52a8\u6001\u4e13\u5bb6\u9009\u62e9<\/strong>\uff1a\u5728\u6bcf\u6b21\u8fed\u4ee3 t \u4e2d\uff0c\u00a0\"\u56fe\u7247\"\u57fa\u4e8e\u524d\u4e00\u6b21\u8fed\u4ee3\u7684\u8f93\u51fa\u00a0\"\u56fe\u7247\"\u00a0\u6765\u51b3\u5b9a\u5f53\u524d\u8fed\u4ee3\u4e2d\u54ea\u4e9b\u4e13\u5bb6\u5e94\u8be5\u88ab\u6fc0\u6d3b\u4ee5\u53ca\u5b83\u4eec\u7684\u6743\u91cd\u3002<\/li>\n
        2. \u4fe1\u606f\u6d41\u63a7\u5236<\/strong>\uff1a\u8def\u7531\u51fd\u6570\u63a7\u5236\u4e86\u4fe1\u606f\u5982\u4f55\u5728\u4e13\u5bb6\u7f51\u7edc\u4e2d\u6d41\u52a8\uff0c\u901a\u8fc7\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u91cd\u65b0\u8bc4\u4f30\u8def\u7531\u51b3\u7b56\uff0c\u786e\u4fdd\u4fe1\u606f\u80fd\u591f\u6d41\u5411\u6700\u76f8\u5173\u7684\u4e13\u5bb6\u3002<\/li>\n
        3. \u9002\u5e94\u6027\u5904\u7406<\/strong>\uff1a\u968f\u7740\u8fed\u4ee3\u7684\u8fdb\u884c\uff0c\u8def\u7531\u51fd\u6570\u53ef\u4ee5\u8c03\u6574\u5176\u8def\u7531\u7b56\u7565\uff0c\u6839\u636e\u5148\u524d\u8fed\u4ee3\u7684\u5904\u7406\u7ed3\u679c\u6765\u4f18\u5316\u540e\u7eed\u4e13\u5bb6\u7684\u9009\u62e9\u3002<\/li>\n<\/ol>\n

          \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u53d1\u73b0<\/h3>\n

          \u6211\u4eec\u91c7\u53d6 DeepSeekV2 \u7ed3\u6784\uff0c\u5728 500M \u7ea7\u522b MoE \u4e0b\u4f7f\u7528 32K Tok \u7684 batch size \u8bad\u7ec3 1000 \u6b65\uff0c\u5b8c\u6210\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\u6765\u9a8c\u8bc1 CoE \u7684\u6709\u6548\u6027\u3002\u5355\u6b21\u8bad\u7ec3\u4f7f\u7528\u4e00\u5f20 H100 \u7ea6\u4e3a 30 \u5206\u949f\uff0c\u4f7f\u7528\u4e00\u5f20 4090 \u7ea6\u4e3a 2 \u5c0f\u65f6\u3002\u66f4\u591a\u5b9e\u9a8c\u7ec6\u8282\u89c1\u540e\u3002<\/p>\n

          \u6211\u4eec\u7684\u5b9e\u9a8c\u65b9\u6cd5\u547d\u540d\u4e3a\"\u56fe\u7247\"\uff0c\u4f8b\u5982 CoE-2 (4\/64) \u4ee3\u8868\u4f7f\u7528 CoE\uff0c\u5faa\u73af\u6b21\u6570\u4e3a 2\uff0c\u6bcf\u6b21\u9009\u53d6 4 \u4e2a\u4e13\u5bb6\uff0c\u6bcf\u5c42\u4e00\u5171\u6709 64 \u4e2a\u4e13\u5bb6\u3002\u6211\u4eec\u5728\u6240\u6709\u5b9e\u9a8c\u4e2d\u8bbe\u7f6e\u5171\u4eab\u4e13\u5bb6\u7684\u6570\u91cf\u4e3a 1\uff0c\u5728\u8fed\u4ee3\u6b21\u6570\u4e3a 2 \u7684\u60c5\u51b5\u4e0b\uff0c\u8fd9\u4f1a\u63d0\u9ad8 5% \u5de6\u53f3\u7684\u8ba1\u7b97\u9700\u6c42\u3002<\/p>\n

          \u6548\u679c\u4f18\u52bf<\/h4>\n

          1\u3001 \u5728\u7b97\u529b\u9884\u7b97\u4e0e\u663e\u5b58\u9884\u7b97\u76f8\u4f3c\u65f6\uff0cCoE \u6548\u679c\u660e\u663e\u66f4\u4f18\u3002<\/strong><\/p>\n

          \u5728 Math \u4efb\u52a1\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3 (\u6b65\u6570 1000)\uff0cCoE-2 (4\/64) \u7684\u6548\u679c\u660e\u663e\u4f18\u4e8e MoE (8\/64)\uff0c\u5728\u76f8\u4f3c\u7684\u7b97\u529b\u548c\u5185\u5b58\u8981\u6c42\u4e0b\uff0c\u5c06 loss \u4ece 1.20 \u4e0b\u964d\u81f3 1.12\uff0c\u4e14\u6709\u66f4\u9661\u5ced\u7684\u4e0b\u964d\u8d8b\u52bf\u3002<\/strong><\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \u6211\u4eec\u8fdb\u4e00\u6b65\u5728 \u201cdense\u201d\uff08\u4e13\u5bb6 8 \u9009 8\uff09\u6a21\u578b\u4e0a\u4e5f\u6d4b\u8bd5\u4e86 CoE\uff0c\u8bc1\u660e\u4e86\u4e32\u884c\u5904\u7406\u5728 Sparse MoE \u4e0a\u76f8\u6bd4 Dense \u6a21\u578b\u66f4\u6709\u6548\uff0cCoE \u662f (Fine-Grained) Sparse MoE \u7684\u4e13\u5c5e\u65b9\u6cd5\u3002\u5982\u56fe\u6240\u793a\uff0c\u91c7\u53d6 2 \u6b21\u5e8f\u5217\u5316\u5904\u7406\u5bf9 Dense \u6a21\u578b\u6027\u80fd\u6ca1\u6709\u663e\u8457\u76ca\u5904\u3002<\/strong><\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \u8d44\u6e90\u6548\u7387<\/h4>\n

          1\u3001\u5728\u8ba1\u7b97\u91cf\u548c\u6548\u679c\u76f8\u4f3c\u7684\u60c5\u51b5\u4e0b\uff0cCoE \u53ef\u4ee5\u51cf\u5c0f\u5bf9\u663e\u5b58\u7684\u8981\u6c42\u3002<\/strong><\/p>\n

          \u4f8b\u5982\uff0cCoE-2 (4\/48) \u7684\u6548\u679c\u4e0e MoE (8\/64) \u76f8\u8fd1\uff0c\u4f46\u4f7f\u7528\u66f4\u5c11\u7684\u603b\u4e13\u5bb6\u6570\u91cf\u3002<\/p>\n

          \u5982\u4e0b\u56fe\u6240\u793a\uff0closs match \u7684\u60c5\u51b5\u4e0b\u51cf\u5c0f\u4e86 17.6% \u7684\u663e\u5b58\u9700\u6c42\u3002<\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \uff08\u6ce8\uff0c\u6211\u4eec\u8bb0\u5f55\u4e86 GPU memory allocated\uff0c\u4e0e micro_bsz=1 \u7684\u65f6\u5019\u7406\u8bba\u6700\u5c0f\u663e\u5b58\u76f8\u5173\uff0c\u5b9e\u9645\u4e0a\u56e0\u4e3a\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u4e86\u8f83\u5927\u7684 batch size\uff0c\u4f7f\u7528\u7684\u663e\u5b58\u4f1a\u66f4\u5927\uff09<\/p>\n

          2\u3001\u5728\u9884\u7b97\u76f8\u4f3c\u7684\u60c5\u51b5\u4e0b\uff0cCoE \u662f\u66f4\u597d\u7684\u8ba1\u7b97\u6269\u5c55\u65b9\u5f0f\u3002<\/strong><\/p>\n

          \u6211\u4eec\u5bf9\u6bd4\u4e86 CoE \u548c\u5176\u4ed6\u8ba1\u7b97\u6269\u5c55\u65b9\u6cd5\uff1a\u62d3\u5c55\u6a21\u578b\u5c42\u6570\u4e0e\u62d3\u5c55\u4e13\u5bb6\u9009\u62e9\u4e2a\u6570\u3002<\/p>\n

          a) \u62d3\u5c55\u8fed\u4ee3\u6b21\u6570 (CoE) > \u62d3\u5c55\u6a21\u578b\u5c42\u6570<\/p>\n

          CoE-2 (8\/64),4 \u5c42 vs MoE (8\/64),8 \u5c42 \/ 12 \u5c42\uff0c8 \u5c42 MoE \u548c CoE \u6548\u679c\u51e0\u4e4e\u76f8\u540c\uff0c\u4f46\u662f\u5bf9 Memory \u8981\u6c42\u9ad8 72%\uff0c\u5373 CoE \u76f8\u5bf9\u8282\u7701\u4e86 42% memory\u3002<\/strong><\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          b) \u62d3\u5c55\u8fed\u4ee3\u6b21\u6570 (CoE)> \u62d3\u5c55\u4e13\u5bb6\u9009\u62e9\u4e2a\u6570<\/p>\n

          CoE-2 (8\/64),4 \u5c42 vs MoE (16\/64) \u6216 MoE (24\/64),4 \u5c42\uff0cMemory \u548c Compute \u8981\u6c42\u4e00\u81f4\uff0cCoE \u6548\u679c\u66f4\u597d\u3002<\/strong><\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          3\u3001\u67b6\u6784\u8bbe\u8ba1\u7684\u5173\u952e\u53d1\u73b0<\/strong><\/p>\n

          a. \u72ec\u7acb\u95e8\u63a7\u673a\u5236<\/p>\n

          \u72ec\u7acb\u95e8\u63a7\u673a\u5236\u53ef\u4ee5\u63d0\u5347\u6a21\u578b\u6027\u80fd\uff0c\u80fd\u4ece\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u89e3\u91ca\u4e13\u5bb6\u7684\u5206\u5316\uff0c\u8868\u660e\u540c\u4e00\u4e13\u5bb6\u5728\u4e0d\u540c\u8fed\u4ee3\u9636\u6bb5\u5904\u7406\u4e0d\u540c\u7c7b\u578b\u7684\u4fe1\u606f\u3002<\/p>\n

          \u6211\u4eec\u5b9e\u73b0\u4e86\u5171\u4eab\u95e8\u63a7\u7684 CoE \u53d8\u79cd\uff0c\u4f7f\u7528\u5171\u4eab\u7684\"\u56fe\u7247\"\u00a0\u800c\u975e\u00a0\"\u56fe\u7247\"\uff0c\u5177\u4f53\u800c\u8a00\uff1a<\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \u53d1\u73b0\u6548\u679c\u5dee\u4e8e CoE\uff0c\u751a\u81f3\u5dee\u4e8e\u6ca1\u6709\u62d3\u5c55\u89c4\u6a21\u7684 MoE\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u4f53\u73b0\u72ec\u7acb\u95e8\u63a7\u673a\u5236\u7684\u91cd\u8981\u6027\u3002<\/p>\n

          b. \u6b8b\u5dee\u8fde\u63a5<\/p>\n

          \u5185\u6b8b\u5dee\u8fde\u63a5 (inner residual) \u6bd4\u5916\u6b8b\u5dee\u8fde\u63a5 (outer residual) \u66f4\u6709\u6548\uff0c\u8fd9\u8868\u660e\u6bcf\u7ec4\u4e13\u5bb6\u5bf9 token \u7684\u4e32\u884c\u5904\u7406\u5b9e\u9645\u4e0a\u662f\u5728\u63d0\u9ad8\u6a21\u578b\u7684\u6709\u6548\u6df1\u5ea6\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u7684\u5904\u7406\u90fd\u5728\u66f4\u597d\u5730\u5b66\u4e60\u6b8b\u5dee\u3002<\/p>\n

          \u6211\u4eec\u5b9e\u73b0\u4e86\u5916\u6b8b\u5dee\u8fde\u63a5\u7684\u53d8\u79cd\uff0c\u5c06\u00a0\"\u56fe\u7247\"\u00a0\u8bbe\u7f6e\u4e3a 0\uff0c\u5c42\u5185\u8fed\u4ee3\u5904\u7406\u7ed3\u675f\u540e\u518d\u8fde\u63a5\u6b8b\u5dee\uff0c\u5177\u4f53\u800c\u8a00\uff1a<\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \u53d1\u73b0\u6548\u679c\u5dee\u4e8e CoE \u4e14\u4e0e MoE \u76f8\u5f53\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u4f53\u73b0\u5185\u6b8b\u5dee\u8fde\u63a5\u7684\u91cd\u8981\u6027\u3002<\/p>\n

          \"\u56fe\u7247\"<\/p>\n

          \u7406\u8bba\u89c2\u5bdf\uff1a\u514d\u8d39\u5348\u9910\u6548\u5e94<\/h4>\n

          \u6700\u663e\u8457\u7684\u662f\uff0cCoE \u6216\u8bb8\u63d0\u4f9b\u4e86\u6211\u4eec\u79f0\u4e4b\u4e3a “\u514d\u8d39\u5348\u9910” \u52a0\u901f\u3002\u901a\u8fc7\u91cd\u6784\u4fe1\u606f\u5728\u6a21\u578b\u4e2d\u7684\u6d41\u52a8\u65b9\u5f0f\uff0c\u4e0e\u4ee5\u5f80 MoE \u65b9\u6cd5\u76f8\u6bd4\uff0c\u6211\u4eec\u4ee5\u66f4\u5c11\u7684\u8ba1\u7b97\u5f00\u9500\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u7ed3\u679c\u3002\u6211\u4eec\u8ba4\u4e3a\u8fd9\u79cd\u6548\u679c\u53ef\u80fd\u6765\u81ea\u4e8e\u4e09\u4e2a\u56e0\u7d20\uff1a<\/p>\n

          1. \u4e13\u5bb6\u9009\u62e9\u7684\u81ea\u7531\u5ea6\u63d0\u9ad8\u3002\"\u56fe\u7247\"\uff0c\u9009\u62e9\u7684\u53ef\u80fd\u6027\u53d8\u591a<\/p>\n

          2. CoE \u7edf\u4e00\u4e86\u4e32\u884c\u5904\u7406 (sequential processing) \u548c\u4e13\u5bb6\u901a\u4fe1 (expert communication) \u4e24\u4e2a\u6982\u5ff5\uff1a<\/p>\n

            \n
          • \u4e0d\u540c\u4e13\u5bb6\u4e4b\u95f4\u53ef\u4ee5\u4e32\u884c\u5904\u7406\uff0c\u63d0\u9ad8\u4e86 Transformer \u7684\u6709\u6548\u6df1\u5ea6\u3002\u4e24\u6b21\u524d\u5411\u4f20\u64ad\u9001\u5230\u4e0d\u540c\u4e13\u5bb6\u65f6\uff0c\u589e\u5f3a\u4e86\u4e13\u5bb6\u7684\u4e32\u884c\u7279\u5f81 \u2014 \u5904\u7406\u540c\u4e00\u4e2a token \u7684\u8def\u7531\u4e13\u5bb6\u6570\u76f8\u540c\uff0c\u4f46\u662f\u662f\u4e32\u884c\u5730\u5904\u7406\u3002<\/li>\n
          • \u4e00\u4e2a\u4e13\u5bb6\u5728\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\u6709\u673a\u4f1a\u5904\u7406\u4e00\u4e2a token \u591a\u6b21\uff0c\u53ef\u80fd\u6709\u52a9\u4e8e\u4fc3\u8fdb\u4e13\u5bb6\u7684\u5206\u5316\uff1a\u4e00\u4e2a\u4e13\u5bb6\u4e0e\u5176\u4ed6\u4e13\u5bb6\u5408\u4f5c\u5904\u7406\u540e\uff0c\u53ef\u4ee5\u5728\u540e\u7eed\u8fed\u4ee3\u5bf9 token \u8fdb\u884c\u8865\u5145\u5904\u7406\u3002<\/li>\n<\/ul>\n

            \u5b9e\u9a8c\u7ec6\u8282<\/h4>\n

            1\u3001\u6570\u636e\u96c6<\/strong><\/p>\n

            \u5b9e\u9a8c\u4f7f\u7528\u7684\u6570\u636e\u96c6\u4e3a MetaMathQA (\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/datasets\/meta-math\/MetaMathQA\u00a0)\uff0c\u8be5\u6570\u636e\u4e3a\u4ece GSM8K \u548c MATH \u6570\u636e\u96c6\u4e2d\u589e\u5f3a\u800c\u6765\uff0c\u4e0d\u542b\u7f51\u7edc\u83b7\u53d6\u7684\u901a\u7528\u6570\u636e\u3002<\/p>\n

            2\u3001\u6a21\u578b\u914d\u7f6e<\/strong><\/p>\n

            \u6211\u4eec\u57fa\u4e8e DeepSeek-V2-Lite \u67b6\u6784\uff0c\u8bbe\u7f6e\u603b\u53c2\u6570\u4e3a 544MB\uff08\u4e0d\u542b embedding\uff09\u7684\u6a21\u578b\uff0c\u5b9e\u73b0\u4e86 CoE \u65b9\u6cd5\uff0c\u4e3b\u8981\u914d\u7f6e\u53c2\u6570\u5982\u4e0b\uff1a<\/p>\n

              \n
            • \u57fa\u7840\u6a21\u578b\u53c2\u6570\uff1a\u9690\u85cf\u5c42\u5927\u5c0f\uff1a1024\uff0c\u9690\u85cf\u5c42\u6570\u91cf\uff1a4\uff0c\u6ce8\u610f\u529b\u5934\u6570\uff1a8<\/li>\n
            • MoE \u76f8\u5173\u53c2\u6570\uff1a\u8def\u7531\u4e13\u5bb6\u603b\u6570\uff1a63\uff0c\u5171\u4eab\u4e13\u5bb6\u6570\u91cf\uff1a1\uff0c\u6bcf\u4e2a token \u9009\u62e9\u7684\u4e13\u5bb6\u6570\u91cf\uff1a8 (routed) + 1 (Shared)\uff0cMoE \u4e2d\u95f4\u5c42\u5927\u5c0f\uff1a704\uff0cMoE \u5c42\u9891\u7387\uff1a1 (\u6bcf\u5c42\u90fd\u662f MoE \u5c42)<\/li>\n
            • CoE \u7279\u6709\u53c2\u6570\uff1a\u8fed\u4ee3\u6b21\u6570\uff1a2\uff0c\u6b8b\u5dee\u8fde\u63a5\u65b9\u5f0f\uff1a\u5185\u90e8\u8fde\u63a5\uff0c\u72ec\u7acb\u95e8\u63a7\u673a\u5236\uff1a\u5f00\u542f (\u6bcf\u6b21\u8fed\u4ee3\u4f7f\u7528\u4e0d\u540c\u7684\u8def\u7531\u51fd\u6570)<\/li>\n<\/ul>\n

              3\u3001\u8bad\u7ec3\u8bbe\u7f6e<\/strong><\/p>\n

                \n
              • \u6279\u91cf\u5927\u5c0f (batch size)\uff1a64\uff0c\u5e8f\u5217\u957f\u5ea6\uff1a512\uff0c\u8bad\u7ec3\u6b65\u6570\uff1a1000\uff0c\u4f18\u5316\u5668\uff1aAdamW\uff0c\u5b66\u4e60\u7387\uff1a3e-4\uff0c\u5b66\u4e60\u7387\u8c03\u5ea6\uff1a10% Warmup\uff0cbetas: [0.9, 0.95]\u00a0 \u00a0weight_decay: 0.01 \u00a0 warmup_steps_ratio: 0.1 \u00a0 clip_grad: 1.0\u3002<\/li>\n<\/ul>\n

                4\u3001\u7cfb\u7edf\u8bbe\u7f6e<\/strong><\/p>\n

                \u8bad\u7ec3\u91c7\u53d6\u6539\u7248\u540e\u7684 veRL \u67b6\u6784\uff08\u5fae\u8c03\u4e86\u539f\u7248\u5b9e\u73b0\uff09\uff0c\u4ee3\u7801\u53c2\u8003\u00a0https:\/\/github.com\/ZihanWang314\/coe\uff0c\u5b9e\u9a8c\u90fd\u5728\u591a\u4e2a\u5355\u5f20 H100 \u7684\u670d\u52a1\u5668\u4e0a\u8fdb\u884c\u3002\u6bcf\u7ec4\u5b9e\u9a8c\u5728 1 \u5c0f\u65f6\u4ee5\u5185\u5b8c\u6210\u3002<\/p>\n

                \u7ed3\u8bba<\/h3>\n

                Chain-of-Experts \u4ee3\u8868\u4e86\u9ad8\u6548\u3001\u9ad8\u6027\u80fd\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u7684\u91cd\u8981\u4e00\u6b65\u3002\u901a\u8fc7\u5728\u7a00\u758f\u795e\u7ecf\u7f51\u7edc\u4e2d\u5b9e\u73b0\u6c9f\u901a\u5904\u7406\uff0cCoE \u89e3\u51b3\u4e86\u5f53\u524d MoE \u67b6\u6784\u4e2d\u7684\u57fa\u672c\u9650\u5236\uff0c\u540c\u65f6\u4ee5\u66f4\u5c11\u7684\u8ba1\u7b97\u9700\u6c42\u63d0\u4f9b\u4f18\u8d8a\u7684\u6027\u80fd\u3002<\/p>\n

                \u8fd9\u9879\u7814\u7a76\u4e3a\u9ad8\u6548\u6269\u5c55\u8bed\u8a00\u6a21\u578b\u5f00\u8f9f\u4e86\u65b0\u9014\u5f84\uff0c\u53ef\u80fd\u4f7f\u5148\u8fdb\u7684\u4eba\u5de5\u667a\u80fd\u80fd\u529b\u66f4\u52a0\u666e\u60e0\u548c\u53ef\u6301\u7eed\u3002<\/p>\n

                \u672a\u6765\u5de5\u4f5c<\/h4>\n

                1. \u6269\u5c55\u6cd5\u5219\u7814\u7a76\uff1a\u63d0\u9ad8\u6a21\u578b\u5c3a\u5bf8\u3001\u9884\u8bad\u7ec3\u6b65\u6570\u3001\u6279\u91cf\u5927\u5c0f\uff0c\u6d4b\u8bd5\u4e0d\u540c\u6570\u636e\u96c6\u7b49\u3002\u76ee\u524d\u4f7f\u7528\u6570\u5b66\u6570\u636e\u96c6\u662f\u56e0\u4e3a\u5176\u63a8\u7406\u590d\u6742\u6027\u5177\u6709\u6311\u6218\u6027\uff0c\u4f46\u8ba1\u5212\u5728\u901a\u7528\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u6d4b\u8bd5\u3002<\/p>\n

                2. \u8fdb\u4e00\u6b65\u8bc4\u4f30\uff1a<\/p>\n

                  \n
                • \u6d4b\u8bd5\u6a21\u578b\u7684\u5b9e\u9645\u4efb\u52a1\u8868\u73b0\u800c\u975e\u4ec5\u9650\u4e8e\u8bc4\u4f30\u635f\u5931<\/li>\n
                • \u6269\u5c55\u5faa\u73af\u6b21\u6570\uff1a\u76ee\u524d\u53ea\u6d4b\u8bd5\u4e86\u5faa\u73af\u6b21\u6570 = 2\uff0c\u63a2\u7d22 3\u30014 \u6b21\u7b49\u662f\u5426\u6709\u6548<\/li>\n
                • \u8fdb\u4e00\u6b65\u8bc4\u4f30\u5171\u4eab\u4e13\u5bb6\u5bf9 CoE \u7684\u4f5c\u7528<\/li>\n<\/ul>\n

                  3. \u67b6\u6784\u521b\u65b0\uff1a<\/p>\n

                    \n
                  • \u57fa\u4e8e\u5faa\u73af\u544a\u8bc9\u6211\u4eec\u4e00\u4e2a\u4e13\u5bb6\u53ef\u4ee5\u4f7f\u7528\u591a\u6b21\u7684\u539f\u7406\uff0c\u63a2\u7d22\u662f\u5426\u53ef\u4ee5\u5b9e\u73b0\u6240\u6709\u5c42\u5171\u4eab\u4e13\u5bb6\u7684 MoE Transformer\uff0c\u5e76\u5206\u5c42\u9009\u62e9\u4e13\u5bb6<\/li>\n
                  • \u4f8b\u5982\uff0cDeepSeekV2Lite \u6709 27 \u5c42\uff0c\u6bcf\u5c42 64 \u4e2a\u4e13\u5bb6\uff0c\u76ee\u524d\u6bcf\u5c42\u9009 6 \u4e2a\u8def\u7531\u4e13\u5bb6\u6709\u7ea6 7\u00d710\u2077\u00a0\u79cd\u7ec4\u5408\uff0c\u800c\u5982\u679c\u6bcf\u4e00\u5c42\u8fc7 MoE \u65f6\uff0c\u90fd\u4ece\u5168\u90e8 27\u00d764\u00a0\u4e2a\u4e13\u5bb6\u4e2d\u9009\u62e9\uff0c\u4e00\u5171\u5c06\u6709 3\u00d710\u00b9\u2076\u00a0\u79cd\u53ef\u80fd\u7684\u7ec4\u5408<\/li>\n<\/ul>\n

                    \u5c40\u9650\u6027<\/h4>\n
                      \n
                    1. \u867d\u7136\u7406\u8bba\u4e0a TFLOPs \u4fdd\u6301\u4e00\u81f4\uff0c\u4f46\u5b9e\u9645\u8bad\u7ec3\u65f6\u95f4\u4f1a\u7565\u5fae\u589e\u52a0\uff0c\u56e0\u4e3a\u5355\u6b21\u8fed\u4ee3\u9009\u62e9\u66f4\u5c11\u4e13\u5bb6\u51cf\u5c0f\u4e86\u77e9\u9635\u4e58\u6cd5\u5e76\u884c\u5ea6\u3002\u672a\u6765\u5de5\u4f5c\u53ef\u80fd\u9700\u8981\u8fdb\u884c\u5e95\u5c42\u9002\u914d\u3002<\/li>\n
                    2. \u8be5\u65b9\u6cd5\u9700\u8981\u4ece\u5934\u9884\u8bad\u7ec3\u800c\u4e0d\u80fd\u7b80\u5355\u5730\u9002\u914d\u73b0\u6709\u6a21\u578b\uff0c\u9650\u5236\u4e86\u5927\u4f17\u7814\u7a76\u8005\u7684\u5e94\u7528\u3002<\/li>\n
                    3. \u6a21\u578b\u66f4\u5927\u65f6\uff0c\u591a\u8282\u70b9 EP \u901a\u4fe1\u5f00\u9500\u4e5f\u9700\u8981\u7eb3\u5165\u8003\u8651\uff1b\u5f53\u524d\u67b6\u6784\u57fa\u4e8e DeepSeekV2\uff0c\u4e0e V3 \u7684\u6a21\u578b\u67b6\u6784\u517c\u5bb9\uff0c\u4f46 V3 \u4e2d\u5b9e\u73b0\u7684 MTP\u3001FP8 \u7b49\u8bad\u7ec3\u67b6\u6784\u5c1a\u672a\u6d4b\u8bd5\uff0c\u517c\u5bb9\u6027\u672a\u77e5\u3002<\/li>\n<\/ol>\n

                      \u8fd9\u7bc7\u535a\u5ba2\u6587\u7ae0\u57fa\u4e8e\u6211\u4eec\u5173\u4e8e Chain-of-Experts (CoE) \u7684\u7814\u7a76\u8bba\u6587\u3002\u6709\u5173\u66f4\u591a\u6280\u672f\u7ec6\u8282\u548c\u5b9e\u9a8c\u7ed3\u679c\uff0c\u8bf7\u5173\u6ce8\u6211\u4eec\u5373\u5c06\u53d1\u5e03\u7684\u5b8c\u6574\u8bba\u6587\u3002<\/p>\n

                      \u8d21\u732e\u8005<\/h3>\n

                      Zihan Wang*<\/p>\n

                      Rui Pan*<\/p>\n

                      Lu Yin*<\/p>\n

                      Manling Li*<\/p>\n

                      Shiwei Liu*<\/p>\n

                      *\u00a0\u5171\u540c\u4f5c\u8005\u3002ZW \u548c SL \u53d1\u8d77\u4e86\u9879\u76ee\u3002LY \u63a2\u7d22\u4e86 MoE \u5185\u5c42\u5904\u7406\u673a\u5236\u3002ZW \u63d0\u51fa\u4e86\u4e32\u884c\u4e13\u5bb6\u548c CoE \u6982\u5ff5\uff0c\u5b9e\u73b0\u4e86\u4ee3\u7801\u5e76\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002SL \u63d0\u51fa\u4e86\u72ec\u7acb\u95e8\u63a7\u673a\u5236\u3002RP \u63a2\u7d22\u4e86\u5c42\u95f4\u91cd\u8981\u6027\u5e76\u8d21\u732e\u4e86\u6570\u636e\u96c6\u3002ML \u8fdb\u884c\u4e86\u5b9e\u9a8c\u6307\u5bfc\u5e76\u63d0\u4f9b\u4e86\u4e91\u8ba1\u7b97\u670d\u52a1\u3002ZW \u548c ML \u51c6\u5907\u4e86\u62a5\u544a\u548c\u53ef\u89c6\u5316\u6750\u6599\u3002<\/p>\n

                      \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n

                      <\/div>\n

                      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

                      <\/div>\n","protected":false},"excerpt":{"rendered":"

                      Chain-of-Experts \u4ee3\u8868\u4e86\u9ad8\u6548\u3001\u9ad8\u6027\u80fd\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u7684\u91cd\u8981\u4e00\u6b65\u3002 \u6211\u4eec\u90fd\u77e5\u9053\uff0cDeepSeek-R […]<\/p>\n

                      <\/div>\n

                      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

                      <\/div>\n","protected":false},"author":56,"featured_media":24874,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,20,80],"tags":[124,820,888],"class_list":["post-24872","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-20","category-80","tag-ai","tag-820","tag-888"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24872","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=24872"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24872\/revisions"}],"predecessor-version":[{"id":24875,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24872\/revisions\/24875"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/24874"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=24872"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=24872"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=24872"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}