{"id":38633,"date":"2025-02-20T00:00:53","date_gmt":"2025-02-19T16:00:53","guid":{"rendered":"https:\/\/17aitech.com\/?p=38633"},"modified":"2025-02-20T00:00:53","modified_gmt":"2025-02-19T16:00:53","slug":"%e6%92%9e%e8%bd%a6deepseek-nsa%ef%bc%8ckimi%e6%9d%a8%e6%a4%8d%e9%ba%9f%e7%bd%b2%e5%90%8d%e7%9a%84%e6%96%b0%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9e%b6%e6%9e%84moba%e5%8f%91%e5%b8%83%ef%bc%8c%e4%bb%a3%e7%a0%81","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=38633","title":{"rendered":"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-19-11\" target=\"_blank\">\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00<\/a><\/p>\n<section>\u6628\u5929\u4e0b\u5348\uff0cDeepSeek \u53d1\u5e03\u4e86\u4e00\u7bc7\u65b0\u8bba\u6587\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u6539\u8fdb\u7248\u7684\u6ce8\u610f\u529b\u673a\u5236 NSA\uff1b\u52a0\u4e0a\u8fd8\u6709\u521b\u59cb\u4eba\u517c CEO \u6881\u6587\u950b\u4eb2\u81ea\u53c2\u4e0e\uff0c\u4e00\u65f6\u4e4b\u95f4\u5438\u5f15\u773c\u7403\u65e0\u6570\uff0c\u53c2\u9605\u62a5\u9053\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650955528&amp;idx=1&amp;sn=2565b3c9d66d2d30c1238831ee9e1e32&amp;scene=21#wechat_redirect\" target=\"_blank\">\u521a\u521a\uff01DeepSeek \u6881\u6587\u950b\u4eb2\u81ea\u6302\u540d\uff0c\u516c\u5f00\u65b0\u6ce8\u610f\u529b\u67b6\u6784 NSA<\/a>\u300b\u3002<\/section>\n<section><\/section>\n<section>\u4f46\u5176\u5b9e\u5c31\u5728\u540c\u4e00\u5929\uff0c\u6708\u4e4b\u6697\u9762\u4e5f\u53d1\u5e03\u4e86\u4e00\u7bc7\u4e3b\u9898\u7c7b\u4f3c\u7684\u8bba\u6587\uff0c\u5e76\u4e14\u540c\u6837\u5de7\u5408\u7684\u662f\uff0c\u6708\u4e4b\u6697\u9762\u521b\u59cb\u4eba\u517c CEO \u6768\u690d\u9e9f\u4e5f\u662f\u8be5\u8bba\u6587\u7684\u7f72\u540d\u4f5c\u8005\u4e4b\u4e00\u3002\u5e76\u4e14\u4e0d\u540c\u4e8e DeepSeek \u53ea\u53d1\u5e03\u4e86\u4e00\u7bc7\u8bba\u6587\uff0c\u6708\u4e4b\u6697\u9762\u8fd8\u53d1\u5e03\u4e86\u76f8\u5173\u4ee3\u7801\u3002\u4e14\u8fd9\u4e9b\u4ee3\u7801\u5df2\u6709\u4e00\u5e74\u7684\u5b9e\u9645\u90e8\u7f72\u9a8c\u8bc1\uff0c\u6709\u6548\u6027\u548c\u7a33\u5065\u6027\u90fd\u6709\u4fdd\u8bc1\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a MoBA \u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5373 Mixture of Block Attention\uff0c\u53ef\u4ee5\u76f4\u8bd1\u4e3a\u300c\u5757\u6ce8\u610f\u529b\u6df7\u5408\u300d\u3002\u636e\u4ecb\u7ecd\uff0cMoBA \u662f\u300c<strong>\u4e00\u79cd\u5c06\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u539f\u7406\u5e94\u7528\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u7684\u521b\u65b0\u65b9\u6cd5\u3002<\/strong>\u300d\u8be5\u65b9\u6cd5\u9075\u5faa\u300c\u66f4\u5c11\u7ed3\u6784\u300d\u539f\u5219\uff0c\u5e76\u4e0d\u4f1a\u5f15\u5165\u9884\u5b9a\u4e49\u7684\u504f\u89c1\uff0c\u800c\u662f\u8ba9\u6a21\u578b\u81ea\u4e3b\u51b3\u5b9a\u5173\u6ce8\u54ea\u4e9b\u4f4d\u7f6e\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aMoBA: Mixture of Block Attention for Long-Context LLMs<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/github.com\/MoonshotAI\/MoBA\/blob\/master\/MoBA_Tech_Report.pdf<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/MoonshotAI\/MoBA<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4e0e DeepSeek \u7684 NSA \u6ce8\u610f\u529b\u673a\u5236\u65b0\u8bba\u6587\u4e00\u6837\uff0c\u6708\u4e4b\u6697\u9762\u8fd9\u7bc7 MoBA \u8bba\u6587\u4e5f\u6536\u83b7\u4e86\u8bf8\u591a\u597d\u8bc4\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-b559be3b81e064fca2a07201e9b71192.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-b559be3b81e064fca2a07201e9b71192.png\"><\/a><\/section>\n<section>\u4e0b\u9762\u6211\u4eec\u5c31\u6765\u89e3\u8bfb\u4e00\u4e0b\u8fd9\u7bc7\u8bba\u6587\u3002<\/section>\n<section><\/section>\n<section>\u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u53d1\u5c55\u5230\u4e86\u524d\u6240\u672a\u6709\u7684\u89c4\u6a21\uff0c\u6211\u4eec\u5e0c\u671b\u5b83\u4eec\u80fd\u591f\u5904\u7406\u6a21\u62df\u4eba\u7c7b\u8ba4\u77e5\u7684\u590d\u6742\u4efb\u52a1\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b9e\u73b0 AGI \u7684\u4e00\u4e2a\u5173\u952e\u80fd\u529b\u662f\u5904\u7406\u3001\u7406\u89e3\u548c\u751f\u6210\u957f\u5e8f\u5217\u7684\u80fd\u529b\uff0c\u8fd9\u5bf9\u4e8e\u4ece\u5386\u53f2\u6570\u636e\u5206\u6790\u5230\u590d\u6742\u63a8\u7406\u548c\u51b3\u7b56\u8fc7\u7a0b\u7684\u5404\u79cd\u5e94\u7528\u81f3\u5173\u91cd\u8981\u3002\u8fd9\u79cd\u5bf9\u6269\u5c55\u4e0a\u4e0b\u6587\u5904\u7406\u80fd\u529b\u7684\u9700\u6c42\u4e0d\u4ec5\u4f53\u73b0\u5728\u957f\u8f93\u5165\u63d0\u793a\u7406\u89e3\u7684\u666e\u53ca\u4e0a\uff0c\u8fd8\u4f53\u73b0\u5728\u6700\u8fd1 Kimi k1.5\u3001DeepSeek-R1 \u548c OpenAI o1\/o3 \u4e2d\u5c55\u793a\u7684\uff0c\u5bf9\u957f\u94fe\u601d\u7ef4\uff08CoT\uff09\u8f93\u51fa\u80fd\u529b\u7684\u63a2\u7d22\u4e2d\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u7531\u4e8e\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u5e26\u6765\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u5448\u4e8c\u6b21\u51fd\u6570\u589e\u957f\uff0c\u6269\u5c55 LLM \u7684\u5e8f\u5217\u957f\u5ea6\u5e76\u975e\u6613\u4e8b\u3002\u8fd9\u4e00\u6311\u6218\u50ac\u751f\u4e86\u4e00\u6ce2\u7814\u7a76\u6d6a\u6f6e\uff0c\u5176\u4e2d\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\u662f\u5229\u7528\u6ce8\u610f\u529b\u5206\u6570\u7684\u56fa\u6709\u7a00\u758f\u6027\u3002\u8fd9\u79cd\u7a00\u758f\u6027\u65e2\u6e90\u4e8e\u6570\u5b66\u4e0a\u7684 softmax \u64cd\u4f5c\uff0c\u4e5f\u6e90\u4e8e\u751f\u7269\u5b66\u4e0a\u7684\u89c2\u5bdf\uff0c\u5373\u4eba\u4eec\u5df2\u5728\u4e0e\u8bb0\u5fc6\u5b58\u50a8\u76f8\u5173\u7684\u5927\u8111\u533a\u57df\u4e2d\u89c2\u5bdf\u5230\u4e86\u7a00\u758f\u8fde\u63a5\u3002<\/section>\n<section><\/section>\n<section>\u73b0\u6709\u65b9\u6cd5\u901a\u5e38\u5229\u7528\u9884\u5b9a\u4e49\u7684\u7ed3\u6784\u7ea6\u675f\uff08\u4f8b\u5982\u57fa\u4e8e sink \u7684\u6ce8\u610f\u529b\u673a\u5236\u6216\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u673a\u5236\uff09\u6765\u5229\u7528\u8fd9\u79cd\u7a00\u758f\u6027\u3002\u867d\u7136\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u80fd\u6709\u6548\uff0c\u4f46\u5b83\u4eec\u5f80\u5f80\u9ad8\u5ea6\u4f9d\u8d56\u4e8e\u7279\u5b9a\u4efb\u52a1\uff0c\u53ef\u80fd\u4f1a\u9650\u5236\u6a21\u578b\u7684\u6574\u4f53\u6cdb\u5316\u80fd\u529b\u3002\u53e6\u4e00\u79cd\u65b9\u6cd5\u662f\u91c7\u7528\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4f8b\u5982 Quest\u3001Minference \u548c RetrievalAttention\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5728\u63a8\u7406\u65f6\u9009\u62e9\u90e8\u5206 token \u8fdb\u884c\u8ba1\u7b97\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u4ee5\u51cf\u5c11\u957f\u5e8f\u5217\u7684\u8ba1\u7b97\u91cf\uff0c\u4f46\u5b83\u4eec\u5e76\u672a\u663e\u8457\u7f13\u89e3\u957f\u4e0a\u4e0b\u6587\u6a21\u578b\u7684\u9ad8\u6602\u8bad\u7ec3\u6210\u672c\u3002<\/section>\n<section><\/section>\n<section>\u6700\u8fd1\uff0c\u7ebf\u6027\u6ce8\u610f\u529b\u6a21\u578b\u4f5c\u4e3a\u4e00\u79cd\u6709\u524d\u666f\u7684\u66ff\u4ee3\u65b9\u6848\u51fa\u73b0\uff0c\u4f8b\u5982 Mamba\u3001RWKV \u548c RetNet\u3002\u8fd9\u4e9b\u65b9\u6cd5\u7528\u7ebf\u6027\u8fd1\u4f3c\u66ff\u4ee3\u4e86\u4f20\u7edf\u7684\u57fa\u4e8e softmax \u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u957f\u5e8f\u5217\u5904\u7406\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u7136\u800c\uff0c\u7531\u4e8e\u7ebf\u6027\u6ce8\u610f\u529b\u4e0e\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u5b58\u5728\u663e\u8457\u5dee\u5f02\uff0c\u9002\u914d\u73b0\u6709\u7684 Transformer \u6a21\u578b\u901a\u5e38\u9700\u8981\u9ad8\u6602\u7684\u8f6c\u6362\u6210\u672c\uff0c\u6216\u8005\u9700\u8981\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u5168\u65b0\u7684\u6a21\u578b\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u4eec\u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\u8bc1\u636e\u4ecd\u7136\u6709\u9650\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u4e00\u4e2a\u5173\u952e\u7684\u7814\u7a76\u95ee\u9898\u968f\u4e4b\u800c\u6765\uff1a\u6211\u4eec\u5982\u4f55\u8bbe\u8ba1\u4e00\u79cd\u7a33\u5065\u4e14\u9002\u5e94\u6027\u5f3a\u7684\u6ce8\u610f\u529b\u67b6\u6784\uff0c\u65e2\u80fd\u4fdd\u7559\u539f\u59cb Transformer \u6846\u67b6\uff0c\u53c8\u80fd\u9075\u5faa\u300c\u66f4\u5c11\u7ed3\u6784\u300d\u7684\u539f\u5219\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u4e0d\u4f9d\u8d56\u9884\u5b9a\u4e49\u504f\u5dee\u7684\u60c5\u51b5\u4e0b\u81ea\u4e3b\u51b3\u5b9a\u5173\u6ce8\u54ea\u4e9b\u90e8\u5206\uff1f\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u8fd9\u79cd\u67b6\u6784\u80fd\u591f\u5728\u5b8c\u5168\u6ce8\u610f\u529b\u548c\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\u4e4b\u95f4\u65e0\u7f1d\u5207\u6362\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u4e0e\u73b0\u6709\u9884\u8bad\u7ec3\u6a21\u578b\u517c\u5bb9\uff0c\u5e76\u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\u548c\u52a0\u901f\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u6b64\u7814\u7a76\u4eba\u5458\u63d0\u51fa\u4e86<strong>\u5757\u6ce8\u610f\u529b\u6df7\u5408\u673a\u5236\uff08Mixture of Block Attention, MoBA\uff09<\/strong>\uff0c\u8fd9\u662f\u4e00\u79cd\u5168\u65b0\u67b6\u6784\uff0c\u5b83\u57fa\u4e8e\u6df7\u5408\u4e13\u5bb6\u7cfb\u7edf\uff08MoE\uff09\u7684\u521b\u65b0\u539f\u7406\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u4e8e Transformer \u6a21\u578b\u7684\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u3002MoE \u6b64\u524d\u4e3b\u8981\u5e94\u7528\u4e8e Transformer \u7684\u524d\u9988\u7f51\u7edc\uff08FFN\uff09\u5c42\uff0c\u800c MoBA \u9996\u6b21\u5c06\u5176\u5f15\u5165\u957f\u4e0a\u4e0b\u6587\u6ce8\u610f\u529b\u673a\u5236\u4e2d\uff0c\u5141\u8bb8\u52a8\u6001\u9009\u62e9\u4e0e\u6bcf\u4e2a\u67e5\u8be2 token \u76f8\u5173\u7684\u5386\u53f2\u5173\u952e\u5757\u548c\u503c\u5757\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u63d0\u9ad8\u4e86 LLM \u7684\u6548\u7387\uff0c\u8fd8\u4f7f\u5176\u80fd\u591f\u5904\u7406\u66f4\u957f\u3001\u66f4\u590d\u6742\u7684\u63d0\u793a\uff0c\u800c\u65e0\u9700\u6309\u6bd4\u4f8b\u589e\u52a0\u8d44\u6e90\u6d88\u8017\u3002<\/section>\n<section><\/section>\n<section>\u76f8\u6bd4\u4e4b\u4e0b\uff0cMoBA \u901a\u8fc7\u5c06\u4e0a\u4e0b\u6587\u5212\u5206\u4e3a\u5757\u5e76\u91c7\u7528\u95e8\u63a7\u673a\u5236\u9009\u62e9\u6027\u5730\u5c06\u67e5\u8be2 token \u8def\u7531\u5230\u6700\u76f8\u5173\u7684\u5757\uff0c\u89e3\u51b3\u4e86\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u6548\u7387\u4f4e\u4e0b\u7684\u95ee\u9898\u3002\u8fd9\u79cd\u5757\u7a00\u758f\u6ce8\u610f\u529b\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u6210\u672c\uff0c\u4e3a\u66f4\u9ad8\u6548\u5730\u5904\u7406\u957f\u5e8f\u5217\u94fa\u5e73\u4e86\u9053\u8def\u3002\u6a21\u578b\u80fd\u591f\u52a8\u6001\u9009\u62e9\u6700\u5177\u4fe1\u606f\u91cf\u7684\u5173\u952e\u5757\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u6027\u80fd\u548c\u6548\u7387\uff0c\u8fd9\u5bf9\u4e8e\u6d89\u53ca\u5927\u91cf\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u4efb\u52a1\u5c24\u4e3a\u6709\u76ca\u3002<\/section>\n<section><\/section>\n<section>\u5728\u8be5\u5de5\u4f5c\u4e2d\uff0c\u6708\u4e4b\u6697\u9762\u8be6\u7ec6\u4ecb\u7ecd\u4e86 MoBA \u7684\u67b6\u6784\uff0c\u9996\u5148\u662f\u5176\u5757\u5212\u5206\u548c\u8def\u7531\u7b56\u7565\uff0c\u5176\u6b21\u662f\u4e0e\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u76f8\u6bd4\u7684\u8ba1\u7b97\u6548\u7387\u3002\u4ed6\u4eec\u4e5f\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u8bc1\u660e\u4e86 MoBA \u5728\u5904\u7406\u957f\u5e8f\u5217\u4efb\u52a1\u4e2d\u7684\u5353\u8d8a\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>MoBA\uff1a\u5c06 MoE \u539f\u7406\u5e94\u7528\u4e8e\u6ce8\u610f\u529b\u673a\u5236<\/strong><\/section>\n<section><\/section>\n<section>MoBA \u901a\u8fc7\u52a8\u6001\u9009\u62e9\u5386\u53f2\u7247\u6bb5\uff08\u5757\uff09\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u6269\u5c55\u4e86 Transformer \u6a21\u578b\u7684\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section>MoBA \u7684\u7075\u611f\u6765\u6e90\u4e8e\u6df7\u5408\u4e13\u5bb6\uff08Mixture of Experts\uff0c\u7b80\u79f0 MoE\uff09\u6280\u672f\u548c\u7a00\u758f\u6ce8\u610f\u529b\u6280\u672f\u3002\u524d\u8005\u4e3b\u8981\u5e94\u7528\u4e8e Transformer \u67b6\u6784\u4e2d\u7684\u524d\u9988\u7f51\u7edc\uff08Feedforward Network\uff0c\u7b80\u79f0 FFN\uff09\u5c42\uff0c\u800c\u540e\u8005\u5df2\u88ab\u5e7f\u6cdb\u7528\u4e8e\u6269\u5c55 Transformer \u4ee5\u5904\u7406\u957f\u6587\u672c\u4e0a\u4e0b\u6587\u3002<\/section>\n<section><\/section>\n<section>\u672c\u6587\u521b\u65b0\u70b9\u5728\u4e8e\u6708\u4e4b\u6697\u9762\u5c06 MoE \u539f\u7406\u5e94\u7528\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u672c\u8eab\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u957f\u5e8f\u5217\u66f4\u9ad8\u6548\u548c\u66f4\u6709\u6548\u7684\u5904\u7406\u3002<\/section>\n<section><\/section>\n<section>MoBA \u4e3b\u8981\u5305\u62ec\u5982\u4e0b\u90e8\u5206\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u53ef\u8bad\u7ec3\u7684\u5757\u7a00\u758f\u6ce8\u610f\u529b<\/strong>\uff1a\u5c06\u5b8c\u6574\u4e0a\u4e0b\u6587\u5206\u5272\u6210\u82e5\u5e72\u5757\uff0c\u6bcf\u4e2a\u67e5\u8be2 token \u90fd\u4f1a\u5b66\u4e60\u5173\u6ce8\u6700\u76f8\u5173\u7684 KV \u5757\uff0c\u4ece\u800c\u5b9e\u73b0\u957f\u5e8f\u5217\u7684\u9ad8\u6548\u5904\u7406\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u65e0\u53c2\u6570\u95e8\u63a7\u673a\u5236<\/strong>\uff1a\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u65e0\u53c2\u6570 top-k \u95e8\u63a7\u673a\u5236\uff0c\u4e3a\u6bcf\u4e2a\u67e5\u8be2 token \u9009\u62e9\u6700\u76f8\u5173\u7684\u5757\uff0c\u786e\u4fdd\u6a21\u578b\u53ea\u5173\u6ce8\u6700\u5177\u4fe1\u606f\u91cf\u7684\u5185\u5bb9\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u5b8c\u5168\u6ce8\u610f\u529b\u4e0e\u7a00\u758f\u6ce8\u610f\u529b\u7684\u65e0\u7f1d\u5207\u6362<\/strong>\uff1aMoBA \u88ab\u8bbe\u8ba1\u4e3a\u5b8c\u5168\u6ce8\u610f\u529b\u673a\u5236\u7684\u7075\u6d3b\u66ff\u4ee3\u65b9\u6848\uff0c\u80fd\u591f\u5728\u5b8c\u5168\u6ce8\u610f\u529b\u548c\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\u4e4b\u95f4\u5b9e\u73b0\u5e73\u6ed1\u8fc7\u6e21\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-10f327bccf4f2a8a667f35ded2829359.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-10f327bccf4f2a8a667f35ded2829359.png\"><\/a><\/section>\n<section>\u603b\u4f53\u800c\u8a00\uff0cMoBA \u6ce8\u610f\u529b\u673a\u5236\u4f7f\u6a21\u578b\u80fd\u591f\u81ea\u9002\u5e94\u4e14\u52a8\u6001\u5730\u5173\u6ce8\u4e0a\u4e0b\u6587\u4e2d\u6700\u6709\u4fe1\u606f\u91cf\u7684\u5757\u3002\u8fd9\u5728\u5904\u7406\u6d89\u53ca\u957f\u6587\u6863\u6216\u5e8f\u5217\u7684\u4efb\u52a1\u65f6\u5c24\u5176\u6709\u76ca\uff0c\u56e0\u4e3a\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e2d\uff0c\u5173\u6ce8\u6574\u4e2a\u4e0a\u4e0b\u6587\u53ef\u80fd\u662f\u4e0d\u5fc5\u8981\u7684\uff0c\u5e76\u4e14\u8ba1\u7b97\u4ee3\u4ef7\u9ad8\u6602\u3002MoBA \u9009\u62e9\u6027\u5730\u5173\u6ce8\u76f8\u5173\u5757\u7684\u80fd\u529b\uff0c\u4f7f\u5f97\u4fe1\u606f\u5904\u7406\u66f4\u52a0\u7ec6\u81f4\u4e14\u9ad8\u6548\u3002<\/section>\n<section><\/section>\n<section>\u6708\u4e4b\u6697\u9762\u901a\u8fc7\u6574\u5408\u6765\u81ea FlashAttention \u548c MoE \u7684\u4f18\u5316\u6280\u672f\uff0c\u5b9e\u73b0\u4e86 MoBA \u7684\u9ad8\u6027\u80fd\u7248\u672c\u3002\u56fe 2 \u5c55\u793a\u4e86 MoBA \u7684\u9ad8\u6548\u7387\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-2f0aacb6048c2821c99c1e6e3caeea97.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-2f0aacb6048c2821c99c1e6e3caeea97.png\"><\/a><\/section>\n<section>MoBA \u5b9e\u73b0\u8fc7\u7a0b\u5305\u542b\u4e94\u4e2a\u6b65\u9aa4\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u6839\u636e\u95e8\u63a7\u7f51\u7edc\u548c\u56e0\u679c\u63a9\u7801\uff0c\u786e\u5b9a\u67e5\u8be2 token \u5bf9 KV \u5757\u7684\u5206\u914d\u3002<\/section>\n<\/li>\n<li>\n<section>\u6839\u636e\u5206\u914d\u7684 KV \u5757\uff0c\u5b89\u6392\u67e5\u8be2 token \u7684\u987a\u5e8f\u3002<\/section>\n<\/li>\n<li>\n<section>\u8ba1\u7b97\u6bcf\u4e2a KV \u5757\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u548c\u5206\u914d\u7ed9\u5b83\u7684\u67e5\u8be2 token\uff0c\u8fd9\u4e00\u6b65\u53ef\u4ee5\u901a\u8fc7\u4e0d\u540c\u957f\u5ea6\u7684 FlashAttention \u6765\u4f18\u5316\u3002<\/section>\n<\/li>\n<li>\n<section>\u5c06\u6ce8\u610f\u529b\u8f93\u51fa\u91cd\u65b0\u6392\u5217\u56de\u5176\u539f\u59cb\u987a\u5e8f\u3002<\/section>\n<\/li>\n<li>\n<section>\u4f7f\u7528\u5728\u7ebf Softmax \u5408\u5e76\u76f8\u5e94\u7684\u6ce8\u610f\u529b\u8f93\u51fa\uff0c\u56e0\u4e3a\u4e00\u4e2a\u67e5\u8be2 token \u53ef\u80fd\u4f1a\u5173\u6ce8\u5176\u5f53\u524d\u5757\u4ee5\u53ca\u591a\u4e2a\u5386\u53f2 KV \u5757\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4e0b\u56fe\u4e3a\u8be5\u7b97\u6cd5\u7684\u5f62\u5f0f\u5316\u63cf\u8ff0\uff0c\u5e76\u5728\u56fe 1b \u4e2d\u8fdb\u884c\u4e86\u53ef\u89c6\u5316\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u57fa\u4e8e MoE \u548c FlashAttention \u5b9e\u73b0 MoBA\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d996510db2a2b7dc1b25f609e5065803.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d996510db2a2b7dc1b25f609e5065803.png\"><\/a><\/section>\n<section><strong>MoBA \u7684\u5b9e\u9a8c\u8868\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u9a8c\u8bc1 MoBA \u7684\u5b9e\u9645\u6548\u679c\uff0c\u6708\u4e4b\u6697\u9762\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217 Scaling Law \u548c\u6d88\u878d\u5b9e\u9a8c\u3002<\/section>\n<section><\/section>\n<section><strong>Full Attention \u4e0e MoBA \u5bf9\u6bd4\u5b9e\u9a8c<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u8bad\u7ec3\u4e86 5 \u4e2a\u6a21\u578b\u6765\u8fdb\u884c\u6bd4\u8f83\uff0c\u8868 1 \u7ed9\u51fa\u8be6\u7ec6\u914d\u7f6e\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-3c4d79c35659f38ce04ff0f954d9641e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-3c4d79c35659f38ce04ff0f954d9641e.png\"><\/a><\/section>\n<section>\u7ed3\u679c\u89c1\u56fe 3a\uff0c\u53ef\u4ee5\u770b\u5230\uff0cFull Attention \u4e0e MoBA \u7684 scaling \u8d8b\u52bf\u975e\u5e38\u76f8\u4f3c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8fd9\u4e24\u79cd\u6ce8\u610f\u529b\u673a\u5236\u4e4b\u95f4\u7684\u9a8c\u8bc1\u635f\u5931\u5dee\u5f02\u5728 1e \u2212 3 \u7684\u8303\u56f4\u5185\u4fdd\u6301\u4e00\u81f4\u3002\u8fd9\u8868\u660e\uff0c\u5c3d\u7ba1 MoBA \u7684\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\u7a00\u758f\u5ea6\u9ad8\u8fbe 75%\uff0c\u4f46\u5b83\u5b9e\u73b0\u4e86\u4e0e\u5b8c\u5168\u6ce8\u610f\u529b\u76f8\u5f53\u7684\u6269\u5c55\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-fe058f3cb13cc92eacfa4775de161861.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-fe058f3cb13cc92eacfa4775de161861.png\"><\/a><\/section>\n<section>\u6b64\u5916\uff0c\u8be5\u56e2\u961f\u4e5f\u9a8c\u8bc1\u4e86 MoBA \u7684\u4e0a\u4e0b\u6587\u53ef\u6269\u5c55\u6027\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u8be5\u56e2\u961f\u5c06\u6700\u5927\u5e8f\u5217\u957f\u5ea6\u4ece 8k \u63d0\u5347\u5230\u4e86 32k\u3002\u7ed3\u679c\u89c1\u56fe 3b\uff0c\u53ef\u89c1\u5c3d\u7ba1\u5728\u6240\u6709\u4e94\u9879\u5b9e\u9a8c\u4e2d\uff0cMoBA \u6700\u540e\u7684\u5757 LM \u635f\u5931\u4e0e Full Attention \u76f8\u6bd4\u7565\u9ad8\uff0c\u4f46\u635f\u5931\u5dee\u8ddd\u6b63\u5728\u9010\u6b65\u7f29\u5c0f\u3002\u8be5\u5b9e\u9a8c\u8868\u660e MoBA \u5177\u6709\u957f\u4e0a\u4e0b\u6587\u53ef\u6269\u5c55\u6027\u3002<\/section>\n<section>\u00a0<\/section>\n<section>\u8be5\u56e2\u961f\u4e5f\u901a\u8fc7\u6d88\u878d\u5b9e\u9a8c\u5bf9 MoBA \u7684\u7ec6\u7c92\u5ea6\u5757\u5206\u5272\u7684\u6709\u6548\u6027\u8fdb\u884c\u4e86\u9a8c\u8bc1\uff0c\u7ed3\u679c\u89c1\u56fe 4\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-42d99604f2ba149b6ff4d4e5ed705aad.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-42d99604f2ba149b6ff4d4e5ed705aad.png\"><\/a><\/section>\n<section>\u53ef\u4ee5\u770b\u5230\uff0cMoBA \u7684\u6027\u80fd\u53d7\u5757\u7c92\u5ea6\u7684\u663e\u8457\u5f71\u54cd\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6700\u7c97\u7c92\u5ea6\u8bbe\u7f6e\uff08\u4ece 8 \u4e2a\u5757\u4e2d\u9009\u62e9 2 \u4e2a\u5757\uff09\u4e0e\u66f4\u7ec6\u7c92\u5ea6\u7684\u8bbe\u7f6e\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u5f02\u4e3a 1e-2\u3002\u8fd9\u4e9b\u53d1\u73b0\u8868\u660e\uff0c\u7ec6\u7c92\u5ea6\u5206\u5272\u4f3c\u4e4e\u662f\u589e\u5f3a MoE \u7cfb\u5217\uff08\u5305\u62ec MoBA\uff09\u6a21\u578b\u6027\u80fd\u7684\u666e\u9002\u6280\u672f\u3002<\/section>\n<section>\u00a0<\/section>\n<section><strong>\u5c06 MoBA \u4e0e Full Attention \u878d\u5408\u5230\u4e00\u8d77\u4f1a\u5982\u4f55\uff1f<\/strong><\/section>\n<section><\/section>\n<section>MoBA \u672c\u5c31\u662f\u4f5c\u4e3a Full Attention \u7684\u7075\u6d3b\u66ff\u4ee3\u800c\u8bbe\u8ba1\u7684\uff0c\u53ef\u4ee5\u6700\u5c0f\u7684\u5f00\u9500\u8f7b\u677e\u5730\u4e92\u76f8\u5207\u6362\uff0c\u5e76\u5b9e\u73b0\u76f8\u5f53\u7684\u957f\u4e0a\u4e0b\u6587\u6027\u80fd\u3002\u6708\u4e4b\u6697\u9762\u5b9e\u9a8c\u8868\u660e\uff0cFull Attention \u548c MoBA \u4e4b\u95f4\u7684\u65e0\u7f1d\u8fc7\u6e21\u53ef\u4ee5\u6210\u4e3a\u9ad8\u6548\u957f\u4e0a\u4e0b\u6587\u9884\u8bad\u7ec3\u7684\u89e3\u51b3\u65b9\u6848\u3002\u7136\u540e\u4ed6\u4eec\u8fd8\u8ba8\u8bba\u4e86\u5206\u5c42\u6df7\u5408\u7b56\u7565\uff0c\u5176\u4f5c\u7528\u4e3b\u8981\u662f\u63d0\u5347\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5b9e\u9a8c\u4e2d\u7684\u6df7\u5408\u8bad\u7ec3\u65f6\uff0c\u6a21\u578b\u4f7f\u7528\u7684\u662f\u4e24\u9636\u6bb5\u65b9\u6848\uff1a\u7b2c\u4e00\u9636\u6bb5\u4f7f\u7528 MoBA \u5b8c\u6210 90% \u7684 token \u8bad\u7ec3\uff0c\u7b2c\u4e8c\u9636\u6bb5\u6a21\u578b\u5207\u6362\u5230 Full Attention \u4f7f\u7528\u5269\u4f59 10% \u7684 token \u8fdb\u884c\u8bad\u7ec3\u3002\u53e6\u5916\uff0c\u5f53\u7136\u4e5f\u4f7f\u7528\u4e86\u7eaf MoBA \u548c\u7eaf Full Attention \u65b9\u6cd5\u4f5c\u4e3a\u5bf9\u6bd4\u3002<\/section>\n<section>\u00a0<\/section>\n<section>\u7ed3\u679c\u89c1\u56fe 5a\uff0c\u4ec5\u4f7f\u7528 MoBA \u65f6\u4f1a\u5bfc\u81f4 trailing token \u7684\u4f4d\u7f6e\u635f\u5931\u66f4\u9ad8\u3002\u91cd\u8981\u7684\u662f\uff0c\u6df7\u5408\u65b9\u6848\u7684\u635f\u5931\u51e0\u4e4e\u4e0e Full Attention \u4e00\u6837\u3002\u8fd9\u4e00\u7ed3\u679c\u51f8\u663e\u4e86\u6df7\u5408\u8bad\u7ec3\u65b9\u6848\u5728\u5e73\u8861\u8bad\u7ec3\u6548\u7387\u548c\u6a21\u578b\u6027\u80fd\u65b9\u9762\u7684\u6709\u6548\u6027\u3002\u66f4\u6709\u8da3\u7684\u662f\uff0c\u5728 MoBA \u548c Full Attention \u4e4b\u95f4\u5207\u6362\u65f6\u6ca1\u6709\u89c2\u5bdf\u5230\u660e\u663e\u7684\u635f\u5931\u5cf0\u503c\uff0c\u8fd9\u518d\u6b21\u8bc1\u660e\u4e86 MoBA \u7684\u7075\u6d3b\u6027\u548c\u7a33\u5065\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-723c7f72e6aa81edbab79a39bfaecb41.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-723c7f72e6aa81edbab79a39bfaecb41.png\"><\/a><\/section>\n<section>\u6708\u4e4b\u6697\u9762\u4e5f\u7814\u7a76\u4e86\u5206\u5c42\u6df7\u5408\u7684\u65b9\u6848\u3002\u8fd9\u79cd\u65b9\u6848\u5f53\u7136\u66f4\u52a0\u590d\u6742\uff0c\u7814\u7a76\u52a8\u673a\u662f\u4ed6\u4eec\u89c2\u5bdf\u5230 MoBA \u6709\u65f6\u4f1a\u5bfc\u81f4 SFT \u671f\u95f4\u7684\u6027\u80fd\u4e0d\u4f73\uff0c\u5982\u56fe 5b \u6240\u793a\u3002<\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u63a8\u6d4b\u8fd9\u53ef\u80fd\u5f52\u56e0\u4e8e SFT \u4e2d\u4f7f\u7528\u7684\u635f\u5931\u63a9\u853d\uff08loss masking\uff09\u2014\u2014 \u63d0\u793a\u8bcd token \u901a\u5e38\u88ab\u6392\u9664\u5728 SFT \u671f\u95f4\u7684\u635f\u5931\u8ba1\u7b97\u4e4b\u5916\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5bf9 MoBA \u7b49\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\u9020\u6210\u7a00\u758f\u68af\u5ea6\u96be\u9898\u3002\u56e0\u4e3a\u5b83\u53ef\u80fd\u4f1a\u963b\u788d\u68af\u5ea6\u7684\u53cd\u5411\u4f20\u64ad\uff0c\u800c\u8fd9\u4e9b\u68af\u5ea6\u6700\u521d\u662f\u6839\u636e\u672a\u63a9\u853d\u7684 token \u8ba1\u7b97\u51fa\u6765\u7684\uff0c\u5e76\u8d2f\u7a7f\u6574\u4e2a\u4e0a\u4e0b\u6587\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u4ed6\u4eec\u63d0\u51fa\u4e86\u53e6\u4e00\u79cd\u6df7\u5408\u65b9\u6cd5 \u2014\u2014 \u5c06\u6700\u540e\u51e0\u5c42 Transformer \u4ece MoBA \u6362\u6210 Full Attention\uff0c\u800c\u5176\u4f59\u5c42\u7ee7\u7eed\u91c7\u7528 MoBA\u3002\u5982\u56fe 5b \u548c\u56fe 5c \u6240\u793a\uff0c\u6b64\u7b56\u7565\u53ef\u4ee5\u663e\u8457\u51cf\u5c11 SFT \u635f\u5931\u3002<\/section>\n<section><\/section>\n<section><strong>\u5927\u8bed\u8a00\u6a21\u578b\u8bc4\u4f30<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f\u57fa\u4e8e Llama 3.1 8B \u6a21\u578b\u5f00\u53d1\u4e86 Llama-8B1M-MoBA\uff0c\u901a\u8fc7\u6301\u7eed\u9884\u8bad\u7ec3\u5c06\u4e0a\u4e0b\u6587\u957f\u5ea6\u4ece 128K \u9010\u6b65\u6269\u5c55\u5230 1M token\u3002\u6a21\u578b\u91c7\u7528\u5757\u5927\u5c0f 4096 \u548c top-K \u53c2\u6570 12 \u7684\u8bbe\u7f6e\uff0c\u5b9e\u73b0\u4e86\u9ad8\u8fbe 95.31% \u7684\u6ce8\u610f\u529b\u7a00\u758f\u5ea6\u3002\u4e3a\u5e73\u8861\u6027\u80fd\uff0c\u4ed6\u4eec\u4fdd\u7559\u6700\u540e\u4e09\u5c42\u4e3a Full Attention\uff0c\u5176\u4f59 29 \u5c42\u4f7f\u7528 MoBA\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-e57e0c7443a6b44bcd4a00237f7b61a3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-e57e0c7443a6b44bcd4a00237f7b61a3.png\"><\/a><\/section>\n<section>\u8bc4\u4f30\u7ed3\u679c\u663e\u793a\uff0cMoBA \u6a21\u578b\u4e0e Full Attention \u6a21\u578b\uff08Llama-8B-1M-Full\uff09\u6027\u80fd\u76f8\u5f53\u3002\u7279\u522b\u662f\u5728\u6700\u5177\u6311\u6218\u6027\u7684 RULER \u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u5373\u4f7f\u5728 62.5% \u7684\u7a00\u758f\u5ea6\u4e0b\uff0cMoBA \u6a21\u578b\u4ecd\u8fbe\u5230\u4e86 0.7818 \u7684\u5f97\u5206\uff0c\u51e0\u4e4e\u4e0e Full Attention \u6a21\u578b\u7684 0.7849 \u6301\u5e73\u3002\u5bf9\u4e8e\u9700\u8981\u5904\u7406 100 \u4e07 token \u7684\u300c\u5927\u6d77\u635e\u9488\u300d\u6d4b\u8bd5\u96c6\u4e2d\uff0c\u6a21\u578b\u540c\u6837\u5c55\u73b0\u51fa\u4f18\u79c0\u7684\u80fd\u529b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-eee74e7303111f0cd587291c555ea9ee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-eee74e7303111f0cd587291c555ea9ee.png\"><\/a><\/section>\n<section><strong>\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027<\/strong><\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u8bc1\u660e\uff0cMoBA \u5728\u4fdd\u6301\u6027\u80fd\u7684\u540c\u65f6\u663e\u8457\u63d0\u5347\u4e86\u6548\u7387\uff1a\u5728\u5904\u7406 1M token \u65f6\uff0c\u6bd4 Full Attention \u6a21\u578b\u5feb 6.5 \u500d\u3002\u5728\u6269\u5c55\u5230 1000 \u4e07 token \u65f6\uff0c\u8f83\u6807\u51c6 Flash Attention \u5b9e\u73b0\u4e86 16 \u500d\u52a0\u901f\uff1b\u901a\u8fc7\u5757\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u548c\u4f18\u5316\u5b9e\u73b0\uff0cMoBA \u5c06\u8ba1\u7b97\u590d\u6742\u5ea6\u4ece\u4e8c\u6b21\u65b9\u964d\u4f4e\u5230\u4e86\u4e9a\u4e8c\u6b21\u65b9\u7ea7\u522b\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4f7f\u5f97 MoBA \u7684\u4f18\u52bf\u5728\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\u65f6\u66f4\u4e3a\u660e\u663e\uff0c\u7279\u522b\u9002\u5408\u5904\u7406\u8d85\u957f\u6587\u672c\u573a\u666f\u3002<\/section>\n<section><\/section>\n<section>\u66f4\u591a\u8be6\u60c5\uff0c\u8bf7\u53c2\u89c1\u539f\u8bba\u6587\u3002<\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u94fe\u63a5\uff1a<\/sup><\/em><\/section>\n<section><em><sup>https:\/\/github.com\/MoonshotAI\/MoBA\/blob\/master\/MoBA_Tech_Report.pdf<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-19-11\" target=\"_blank\">\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u649e\u8f66DeepSeek  [&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":[82],"class_list":["post-38633","post","type-post","status-publish","format-standard","hentry","category-news","tag-82"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00 - \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=38633\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=38633\",\"url\":\"https:\/\/17aitech.com\/?p=38633\",\"name\":\"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=38633#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=38633#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png\",\"datePublished\":\"2025-02-19T16:00:53+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=38633#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=38633\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=38633#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=38633#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00\"}]},{\"@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":"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00 - \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=38633","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=38633","url":"https:\/\/17aitech.com\/?p=38633","name":"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=38633#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=38633#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png","datePublished":"2025-02-19T16:00:53+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=38633#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=38633"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=38633#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1cc05bcf7bea2cfa5b2f67ae70923f5d.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=38633#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u649e\u8f66DeepSeek NSA\uff0cKimi\u6768\u690d\u9e9f\u7f72\u540d\u7684\u65b0\u6ce8\u610f\u529b\u67b6\u6784MoBA\u53d1\u5e03\uff0c\u4ee3\u7801\u4e5f\u516c\u5f00"}]},{"@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\/38633","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=38633"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/38633\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38633"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38633"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38633"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}