{"id":50934,"date":"2024-06-16T08:34:35","date_gmt":"2024-06-16T00:34:35","guid":{"rendered":"https:\/\/swarma.org\/?p=50934"},"modified":"2024-06-16T08:34:35","modified_gmt":"2024-06-16T00:34:35","slug":"%e6%8f%90%e9%ab%98%e5%a4%a7%e6%a8%a1%e5%9e%8b%e8%ae%a1%e7%ae%97%e6%95%88%e7%8e%87%ef%bc%81%e5%bd%a9%e4%ba%91%e7%a7%91%e6%8a%80%e6%8e%a8%e5%87%ba%e5%a4%a7%e6%a8%a1%e5%9e%8b-dcformer%ef%bc%8c%e6%95%88","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=50934","title":{"rendered":"\u63d0\u9ad8\u5927\u6a21\u578b\u8ba1\u7b97\u6548\u7387\uff01\u5f69\u4e91\u79d1\u6280\u63a8\u51fa\u5927\u6a21\u578b DCFormer\uff0c\u6548\u7387\u662f Transformer \u4e24\u500d"},"content":{"rendered":"
\n
\n
\n
\n
<\/div>\n<\/section>\n
\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\u6700\u8fd1KAN\u7a81\u7136\u7206\u706b\uff0c\u8ba9\u4eba\u4eec\u5f00\u59cb\u610f\u8bc6\u5230\u5927\u6a21\u578b\u7684\u8ba1\u7b97\u6548\u7387\u81f3\u5173\u91cd\u8981\uff0c\u800c\u63d0\u9ad8\u5927\u578b\u6a21\u578b\u751f\u6210tokens\u7684\u901f\u5ea6\u540c\u6837\u81f3\u5173\u91cd\u8981\u3002\u76f8\u5bf9\u4e8e\u4e0d\u65ad\u52a0\u7801GPU\uff0c\u6539\u5584Transformer\u6a21\u578b\u67b6\u6784\u7684\u8ba1\u7b97\u6548\u7387\u662f\u4e00\u79cd\u66f4\u4e3a\u957f\u8fdc\u6709\u6548\u7684\u65b9\u6cd5\u3002\u8fd1\u671f\uff0c\u5f69\u4e91\u79d1\u6280\u7684\u7814\u7a76\u56e2\u961f\u5bf9Transformer\u8ba1\u7b97\u6700\u8017\u65f6\u7684\u6838\u5fc3\u7ec4\u4ef6\u2014\u2014\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\uff08MHA\uff09\u8fdb\u884c\u4e86\u6539\u8fdb\uff0c\u63a8\u51fa\u52a8\u6001\u7ec4\u5408\u591a\u5934\u6ce8\u610f\u529b\u6539\u8fdbTransformer\uff08DCFormer\uff09\uff0c\u5c06Transformer\u7684\u8ba1\u7b97\u6027\u80fd\u63d0\u5347\u67092\u500d\u4e4b\u9ad8\u3002\u8be5\u8bba\u6587\u5df2\u88abICML 2024\u63a5\u6536\u4e3aoral\uff08\u4eca\u5e74oral\u7684\u63a5\u6536\u7387\u4ec5\u4e3a1.5%\uff09\u3002<\/span><\/strong><\/section>\n

<\/span><\/strong><\/section>\n
\u96c6\u667a\u4ff1\u4e50\u90e8\u9080\u8bf7\u8bba\u6587\u4e00\u4f5c\u3001\u5f69\u4e91\u79d1\u6280\u9996\u5e2d\u79d1\u5b66\u5bb6\u8096\u8fbe\u8001\u5e08\u5728<\/span><\/strong>\u300c<\/span><\/strong>\u540eChatGPT\u65f6\u4ee3\u300d\u8bfb\u4e66\u4f1a<\/span><\/strong><\/a>\u4e2d\u6df1\u5165\u89e3\u8bfb\u4e86DCFormer\u7684\u4e3b\u8981\u5de5\u4f5c\uff0c\u4ecb\u7ecd\u8fd9\u9879\u5de5\u4f5c\u80cc\u540e\u7684\u52a8\u673a\u548c\u7814\u7a76\u5386\u7a0b\uff0c\u5e76\u63a2\u8ba8\u6a21\u578b\u7ed3\u6784\u7814\u7a76\u7684\u672a\u6765\u53d1\u5c55\u65b9\u5411\uff0c\u672c\u6587\u662f\u5bf9\u6b64\u6b21\u8bfb\u4e66\u4f1a\u7684\u6574\u7406\u3002DCFomer\u7684\u8bba\u6587\u4f5c\u8005\u5747\u6765\u81ea\u5f69\u4e91\u79d1\u6280NLP\u7b97\u6cd5\u7ec4\uff0c\u6838\u5fc3\u6210\u5458\u51fa\u81ea\u96c6\u667a\u4ff1\u4e50\u90e8\u3002\u5f69\u4e91\u79d1\u6280\u65d7\u4e0b\u4ea7\u54c1\u5305\u62ec\u5f69\u4e91\u5929\u6c14\u3001\u5f69\u4e91\u5c0f\u68a6\u3001\u5f69\u4e91\u5c0f\u8bd1\u7b49\uff0c\u81f4\u529b\u4e8e\u901a\u8fc7\u4eba\u5de5\u667a\u80fd\u4ea7\u54c1\u8ba9\u751f\u6d3b\u66f4\u7f8e\u597d\u3002\u56e2\u961f\u76ee\u524d\u6b63\u5728\u62db\u8058\u591a\u4e2a\u7b97\u6cd5\u5c97\u4f4d\uff0c\u671f\u5f85\u5fd7\u540c\u9053\u5408\u7684\u670b\u53cb\u52a0\u5165\uff01\u8be6\u60c5\u89c1\u6587\u672b\u3002<\/span><\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/strong><\/span><\/strong><\/section>\n

<\/strong><\/strong><\/span><\/strong><\/section>\n
\u7814\u7a76\u9886\u57df\uff1a\u5927\u6a21\u578b\uff0cTransformer\u67b6\u6784\uff0c\u52a8\u6001\u7ec4\u5408\u591a\u5934\u6ce8\u610f\u529b<\/span><\/strong><\/strong><\/strong><\/span><\/strong>
<\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
 <\/span>\u8096\u8fbe<\/span><\/span> <\/span>| \u8bb2\u8005<\/span><\/strong><\/section>\n
Mugicaxu<\/span><\/span> <\/span>| \u6574\u7406<\/span><\/strong><\/span><\/strong><\/section>\n

<\/span><\/strong><\/section>\n

<\/span><\/section>\n

<\/p>\n

\n
\u8bba\u6587\u6807\u9898\uff1aImproving Transformers with Dynamically Composable Multi-Head Attention<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2405.08553<\/span><\/section>\n
\u5f00\u6e90\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/Caiyun-AI\/DCFormer<\/span><\/section>\n<\/blockquote>\n

<\/span><\/section>\n
\u5728\u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u65e0\u8bba\u662f\u5904\u7406\u6587\u672c\u3001\u56fe\u50cf\u3001\u8bed\u97f3\u8fd8\u662f\u89c6\u9891\u4efb\u52a1\uff0c\u5927\u591a\u4ee5Transformer\u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\u3002\u524d\u6bb5\u65f6\u95f4\u4e00\u7bc7\u975e\u5e38\u53d7\u6b22\u8fce\u7684\u8bba\u6587\uff0c\u540d\u4e3aKAN\uff0c\u5b83\u4e4b\u6240\u4ee5\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\uff0c\u662f\u56e0\u4e3a\u4eba\u4eec\u8ba4\u4e3aKAN\u53ef\u80fd\u4f1a\u5bf9Transformer\u8fdb\u884c\u6539\u8fdb\uff0c\u63d0\u5347\u6a21\u578b\u7684\u6548\u7387\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u540c\u6837\u65e8\u5728\u6539\u8fdbTransformer\u3002<\/span>\uff08\u53c2\u770b\u300aKAN\u4e00\u4f5c\u5218\u5b50\u9e23\u76f4\u64ad\u603b\u7ed3\uff1aKAN\u7684\u80fd\u529b\u8fb9\u754c\u548c\u5f85\u89e3\u51b3\u7684\u95ee\u9898<\/a>\u300b\uff09<\/span><\/section>\n

<\/span><\/section>\n
\u6211\u628aTransformer\u5f15\u9886\u7684\u65b0\u4e00\u4ee3\u6a21\u578b\u8303\u5f0f\u6bd4\u4f5c\u642d\u79ef\u6728\u7684\u6e38\u620f\uff0c\u5e76\u5c06\u5176\u5206\u4e3a\u4e09\u4e2a\u6d3e\u522b\uff1a\u5927\u529b\u98de\u7816\u6d3e\u3001\u7cfb\u7edf\u4ea4\u4e92\u6d3e\u548c\u7ed3\u6784\u521b\u65b0\u6d3e\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n

<\/span><\/section>\n
\u9996\u5148\u662f\u201c\u5927\u529b\u98de\u7816\u6d3e\u201d\uff0c\u4e5f\u5c31\u662f\u6211\u4eec\u5e38\u8bf4\u7684\u201c\u5927\u529b\u51fa\u5947\u8ff9\u201d\u7684 Scaling Law\u3002\u5982\u679c\u6211\u4eec\u628aTransformer block\u770b\u4f5c\u79ef\u6728\u5757\uff0c\u90a3\u4e48\u4f7f\u7528\u66f4\u591a\u7684\u79ef\u6728\u5757\u548c\u6570\u636e\u5c31\u80fd\u5f97\u5230\u66f4\u597d\u7684\u6548\u679c\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u6700\u5927\u4f18\u52bf\u5728\u4e8e\u786e\u5b9a\u6027\u5f3a\uff0c\u56e0\u6b64\u6211\u4eec\u79f0\u4e4b\u4e3a\u201c\u9760\u8c31\u9752\u5e74\u201d\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u53e6\u4e00\u6d3e\u2014\u2014\u201c\u7cfb\u7edf\u4ea4\u4e92\u6d3e\u201d\u8ba4\u4e3a\uff0c\u4e0d\u5e94\u5c06\u6a21\u578b\u89c6\u4e3a\u5b64\u7acb\u7684\u70b9\uff0c\u800c\u5e94\u5c06\u5176\u7f6e\u4e8e\u7cfb\u7edf\u4e2d\uff0c\u8ba9\u5176\u8fdb\u884c\u66f4\u4e30\u5bcc\u7684\u4ea4\u4e92\uff0c\u4f8b\u5982\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u6216\u63d0\u4f9b\u601d\u7ef4\u94fe\u63d0\u793a\u7b49\u3002\u8fd9\u79cd\u601d\u8def\u66f4\u63a5\u8fd1\u4eba\u7c7b\u7684\u5b66\u4e60\u65b9\u5f0f\uff0c\u56e0\u6b64\u6211\u4eec\u79f0\u4e4b\u4e3a\u201c\u6587\u827a\u9752\u5e74\u201d\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u7b2c\u4e09\u4e2a\u6d3e\u522b\u2014\u2014\u201c\u7ed3\u6784\u521b\u65b0\u6d3e\u201d\u5219\u66f4\u4e3a\u76f4\u63a5\uff0c\u4ed6\u4eec\u8ba4\u4e3a\uff0c\u65e2\u7136\u524d\u4e24\u6d3e\u90fd\u628aTransformer\u7684\u79ef\u6728\u5757\u5f53\u4f5c\u9ed1\u76d2\u6765\u4f7f\u7528\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u62c6\u5f00\u770b\u770b\u91cc\u9762\u7a76\u7adf\u662f\u4ec0\u4e48\u5462\uff1f\u4ed6\u4eec\u4e3b\u5f20\u6df1\u5165\u7814\u7a76Transformer\u7684\u5de5\u4f5c\u539f\u7406\uff0c\u63a2\u8ba8\u5982\u4f55\u6539\u8fdb\u751a\u81f3\u66ff\u6362\u5b83\u3002\u8fd9\u4e00\u6d3e\u7684\u505a\u6cd5\u786e\u5b9e\u5f15\u53d1\u4e86\u4e00\u4e9b\u8d28\u7591\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u8bb8\u591a\u4eba\u4f1a\u95ee\uff1a\u201c\u771f\u7684\u6709\u5fc5\u8981\u6539\u8fdbTransformer\u5417\uff1f\u8fd9\u4e2a\u7ed3\u6784\u4e0d\u662f\u5df2\u7ecf\u8db3\u591f\u597d\u4e86\u5417\uff1f\u201d\u8fd9\u4e9b\u7591\u95ee\u5e76\u975e\u65e0\u7684\u653e\u77e2\uff0c\u56e0\u4e3aTransformer\u5df2\u7ecf\u8bde\u751f\u4e03\u5e74\u4e86\uff0c\u65e0\u6570\u5b9e\u8df5\u4f3c\u4e4e\u5df2\u7ecf\u8bc1\u660e\u5b83\u7684\u5f3a\u5927\u548c\u901a\u7528\u6027\u3002\u5e76\u4e14\u6709\u8bba\u6587\u663e\u793a\u5bf9Transformer\u7684\u6539\u8fdb\u5927\u591a\u5e76\u4e0d\u50cf\u539f\u59cbTransformer\u90a3\u6837\u666e\u9002\u6709\u6548\u548c\u53ef\u6269\u5c55\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u6b63\u662f\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\uff0c\u6211\u4eec\u51b3\u5b9a\u8fdb\u884c\u8fd9\u9879\u770b\u4f3c\u6709\u4e9b\u4e2d\u4e8c\u7684\u5de5\u4f5c\u2014\u2014\u6539\u8fdbTransformer\u7684\u7ed3\u6784\u3002\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u540d\u4e3a\u52a8\u6001\u7ec4\u5408\u591a\u5934\u6ce8\u610f\u529b<\/strong><\/span>\uff08DCMHA\uff09<\/span>\u7684\u65b0\u67b6\u6784\uff0c\u7528\u4ee5\u66ff\u6362Transformer\u4e2d\u7684\u591a\u5934\u6ce8\u610f\u529b<\/span>\uff08MHA\uff09<\/span>\u673a\u5236\uff0c\u4ece\u800c\u5f97\u5230\u4e00\u4e2a\u5177\u6709\u66f4\u9ad8\u8ba1\u7b97\u6548\u7387\u7684\u6a21\u578b\u3002\u8fd9\u610f\u5473\u7740\u5728\u8fbe\u5230\u76f8\u540c\u6548\u679c\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u53ef\u4ee5\u8282\u77011.7\u52302\u500d\u7684\u7b97\u529b\uff0c\u6216\u8005\u5728\u4f7f\u7528\u76f8\u540c\u7b97\u529b\u7684\u60c5\u51b5\u4e0b\u83b7\u5f97\u66f4\u597d\u7684\u6548\u679c<\/strong>\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u8fd9\u4e2a\u65b0\u6a21\u578b\u8ba9\u6211\u4eec\u80fd\u591f\u4ee5\u66f4\u5c11\u7684\u7b97\u529b\u8bad\u7ec3\u51fa\u66f4\u4f18\u79c0\u7684\u6a21\u578b\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u672c\u6b21\u5206\u4eab\u56f4\u7ed5\u4e09\u4e2a\u65b9\u9762\u5c55\u5f00\uff1a\u7b2c\u4e00\u90e8\u5206\u662f\u6539\u8fdb\u7684\u52a8\u673a\uff0c\u5373\u6211\u4eec\u662f\u5982\u4f55\u4ea7\u751f\u8fd9\u4e2a\u6539\u8fdb\u60f3\u6cd5\u7684\uff1b\u7b2c\u4e8c\u90e8\u5206\u662fDCFormer\u672c\u8eab\u7684\u8bbe\u8ba1\u548c\u5b9e\u73b0\u7ec6\u8282\uff1b\u7b2c\u4e09\u90e8\u5206\u662f\u5bf9DCFormer\u7814\u7a76\u5386\u7a0b\u7684\u56de\u987e\uff0c\u5305\u62ec\u6211\u4eec\u7684\u7ecf\u9a8c\u6559\u8bad\u548c\u5bf9\u672a\u6765\u7684\u5c55\u671b\uff0c\u63a2\u8ba8\u7ed3\u6784\u6539\u8fdb\u548c\u521b\u65b0\u7684\u53ef\u80fd\u65b9\u5411\u3002<\/span><\/section>\n

<\/span><\/section>\n

<\/span><\/section>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

1. \u52a8\u673a\uff1a\u591a\u5934\u6ce8\u610f\u529b\u7684\u5de5\u4f5c\u673a\u5236\u548c\u95ee\u9898\u5206\u6790<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

 
<\/p>\n

Transformer \u7ec4\u4ef6\uff1aMHA \u548c MLP<\/span><\/strong><\/span><\/h5>\n

<\/h5>\n
    \n
  • \n
    \u591a\u5934\u6ce8\u610f\u529b\uff08MHA\uff09\uff1a\u6838\u5fc3\u7ec4\u4ef6\uff0c\u8de8token\u4fe1\u606f\u6d41\u52a8<\/span><\/strong><\/span><\/h5>\n<\/li>\n<\/ul>\n

    <\/span><\/section>\n
    \u6211\u4eec\u90fd\u77e5\u9053Transformer\u7531\u4e24\u5927\u6a21\u5757\u7ec4\u6210\uff0c\u4e00\u5757\u53eb\u505a\u591a\u5934\u6ce8\u610f\u529b<\/span>\uff08MHA\uff09<\/span>\uff0c\u53e6\u4e00\u5757\u53eb\u505a\u5168\u8fde\u63a5\u5c42<\/span>\uff08MLP\uff09<\/span>\u3002\u524d\u4e00\u6bb5\u6bd4\u8f83\u5f88\u706b\u7684KAN\u67b6\u6784\u672c\u8d28\u4e0a\u662f\u4e00\u9879\u5173\u4e8eMLP\u7684\u5de5\u4f5c\uff0c\u800cDCFormer\u5219\u66f4\u52a0\u5173\u6ce8\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757<\/span>\uff08MHA\uff09<\/span>\u3002\u6211\u4eec\u8ba4\u4e3a\uff0c\u591a\u5934\u6ce8\u610f\u529b\u7684\u91cd\u8981\u6027\u5728\u4e8e\uff1aTransformers\u4e2d\u8de8token\u7684\u4fe1\u606f\u6d41\u52a8\u90fd\u662f\u5728\u6ce8\u610f\u529b\u6a21\u5757\u4e2d\u5b8c\u6210\u7684\uff0c\u800cMLP\u6a21\u5757\u4e2d\u5219\u6ca1\u6709\u8fd9\u79cd\u4fe1\u606f\u6d41\u52a8\u3002\u56e0\u6b64\uff0c\u5f88\u663e\u7136\uff0c\u6240\u6709\u5173\u4e8e\u4e0a\u4e0b\u6587\u7684\u4fe1\u606f\u7684\u6574\u5408\u5904\u7406\u4ee5\u53ca\u7b97\u6cd5\u7684\u5b9e\u73b0\u90fd\u8981\u4f9d\u9760Attention\u6ce8\u610f\u529b\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/section>\n
    \u6211\u4eec\u7684\u7b2c\u4e00\u6b65\u5de5\u4f5c\u662f\u7814\u7a76\u6ce8\u610f\u529b\u6a21\u5757\u91cc\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u662f\u5982\u4f55\u5de5\u4f5c\u7684\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u9700\u8981\u5f15\u5165\u4e00\u4e2a\u56de\u8def<\/span>\uff08Transformer circuits\uff09<\/span>\u7684\u6982\u5ff5\u3002\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u6709\u56db\u4e2a\u6743\u91cd\u77e9\u9635\uff1aWq<\/sub>\u3001Wk<\/sub>\u3001Wo<\/sub>\u3001Wv<\/sub>\uff0c\u5f53\u5b83\u4eec\u4e24\u4e24\u8fdb\u884c\u914d\u5bf9\u65f6\uff0c\u4fbf\u4f1a\u5f62\u6210\u4e86\u4e24\u4e2a\u4fe1\u606f\u56de\u8def\u2014\u2014QK\u548cOV\u3002QK\u56de\u8def\u51b3\u5b9a\u4e86\u4ece\u5f53\u524d\u4f4d\u7f6e\u5173\u6ce8\u4e0a\u4e0b\u6587\u4e2d\u7684\u54ea\u4e9b\u4f4d\u7f6e<\/span>\uff08\u5982\u4e0b\u56fe\u7d2b\u8272\u90e8\u5206\u6240\u793a\uff09<\/span>\u3002OV\u56de\u8def\u51b3\u5b9a\u4e86\u4ece\u5173\u6ce8\u5230\u7684\u4f4d\u7f6e\u53d6\u56de\u4ec0\u4e48\u4fe1\u606f\/\u5c5e\u6027\u5199\u5165\u5f53\u524d\u4f4d\u7f6e\u7684\u6b8b\u5dee\u6d41\uff0c\u8fdb\u800c\u9884\u6d4b\u4e0b\u4e00\u4e2atoken<\/span>\uff08\u5982\u4e0b\u56fe\u9ec4\u8272\u90e8\u5206\u6240\u793a\uff09<\/span>\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/span><\/section>\n

    <\/span><\/section>\n
    Transformer circuits [by Anthropic]<\/span><\/strong><\/span><\/h5>\n

    <\/span><\/strong><\/section>\n
    \n
    Nelson Elhage, et al. A Mathematical Framework for Transformer Circuits. https:\/\/transformer-circuits.pub\/2021\/framework\/index.html<\/span><\/section>\n<\/blockquote>\n

    <\/span><\/section>\n
      \u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684Wq<\/sub>\u3001Wk<\/sub>\u3001Wo<\/sub>\u3001Wv<\/sub>\u5f62\u6210\u4e24\u4e2a\u56de\u8def<\/span><\/section>\n
     \u2022 QK\/\u67e5\u627e\u56de\u8def\uff1a\u51b3\u5b9a\u4ece\u5f53\u524d\u4f4d\u7f6e\u5173\u6ce8\u4e0a\u4e0b\u6587\u4e2d\u7684\u54ea\u4e9b\u4f4d\u7f6e<\/span><\/section>\n

     \u2022 OV\/\u53d8\u6362\u56de\u8def\uff1a\u51b3\u5b9a\u4ece\u5173\u6ce8\u5230\u7684\u4f4d\u7f6e\u53d6\u56de\u4ec0\u4e48\u4fe1\u606f\/\u5c5e\u6027\u5199\u5165\u5f53\u524d\u4f4d\u7f6e\u7684\u6b8b\u5dee\u6d41\uff0c\u8fdb\u800c\u9884\u6d4b\u4e0b\u4e00\u4e2atoken<\/span><\/p>\n


    <\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u56fe\u4e2d\u5c55\u793a\u4e86\u4e00\u4e2a\u7ecf\u5178\u6848\u4f8b\u2014\u2014\u6211\u4eec\u53ef\u4ee5\u79f0\u4e4b\u4e3a\u201c\u5f52\u7eb3\u5934\u201d<\/span>\uff08induction head\uff09<\/span>\u3002\u5bf9\u4e8e\u81f4\u529b\u4e8e\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u7684\u7814\u7a76\u8005\u6765\u8bf4\uff0c\u7406\u89e3\u8fd9\u79cd\u6a21\u578b\u901a\u8fc7\u5185\u90e8\u56de\u8def\u5b8c\u6210\u4efb\u52a1\u7684\u6982\u5ff5\u81f3\u5173\u91cd\u8981\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u901a\u8fc7\u4e00\u4e9b\u5177\u4f53\u7684\u4f8b\u5b50\u6765\u63a2\u8ba8\u8fd9\u79cd\u7ec4\u5408\u65b9\u5f0f\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u4ee5\u53ca\u5b83\u53ef\u80fd\u9762\u4e34\u7684\u95ee\u9898\u3002\u4ee5\u4e2d\u6587\u53e5\u5b50\u201c\u5c0f\u660e\u6709\u4e2a\u94a2\u7b14 … \u5c0f\u660e\u6709\u4e2a -> \u94a2\u7b14\u201d\u4e3a\u4f8b\uff0c\u6a21\u578b\u4f1a\u4f7f\u7528\u4e00\u4e2a\u7c7b\u4f3cinduction head\u7684\u673a\u5236\u6765\u786e\u5b9a\u5982\u4f55\u4ece\u524d\u9762\u7684\u5185\u5bb9\u4e2d\u63d0\u53d6\u4fe1\u606f\u3002\u6a21\u578b\u4f1a\u8bc6\u522b\u5230\u201c\u5c0f\u660e\u201d\u4f5c\u4e3a\u4e3b\u8bed\uff0c\u5e76\u4e0e\u5bbe\u8bed\u201c\u94a2\u7b14\u201d\u6709\u5173\u8054\u3002\u901a\u8fc7\u8bc6\u522b\u524d\u6587\u76f8\u540c\u7684\u4e3b\u8bed\uff0c\u6a21\u578b\u7528\u4e00\u4e2aQK\u56de\u8def\u80fd\u591f\u627e\u5230\u76f8\u5173\u7684\u4fe1\u606f\u2014\u2014\u8fd9\u91cc\u662f\u201c\u94a2\u7b14\u201d\uff0c\u5e76\u7528\u4e00\u4e2aOV\u56de\u8def\u76f4\u63a5\u590d\u5236\u4e86\u201c\u94a2\u7b14\u201d\u8fd9\u4e2a\u4fe1\u606f\u5230\u6b8b\u5dee\u6d41\u7528\u4e8e\u8f93\u51fa\u9884\u6d4b\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u518d\u770b\u4e24\u4e2a\u4f8b\u5b50\uff1a\u201c\u82f9\u679c\u662f\u4e00\u79cd -> \u6c34\u679c\u3002\u6a21\u578b\u4f7f\u7528QK\u6765\u5173\u6ce8\u524d\u9762\u7684\u8bcd\u6c47\u201c\u82f9\u679c\u201d\uff0c\u5e76\u901a\u8fc7OV\u56de\u8def\uff0c\u6765\u63d0\u53d6\u201c\u82f9\u679c\u201d\u7684\u7c7b\u522b\u4fe1\u606f\uff0c\u5373\u201c\u6c34\u679c\u201d\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u53e6\u4e00\u4e2a\u4f8b\u5b50\u662f\uff1a\u201c\u94a2\u7b14\u7528\u6765 -> \u5199\u5b57\u3002\u201d\u8fd9\u91cc\uff0c\u6a21\u578b\u4f7f\u7528QK\u6765\u8bc6\u522b\u201c\u94a2\u7b14\u201d\uff0c\u5e76\u901a\u8fc7OV\u6765\u63d0\u53d6\u5173\u4e8e\u201c\u94a2\u7b14\u201d\u7684\u7528\u9014\u4fe1\u606f\uff0c\u5373\u5b83\u662f\u7528\u6765\u5199\u5b57\u7684\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u5c31\u80fd\u591f\u5229\u7528\u8fd9\u4e9b\u4fe1\u606f\u6765\u8fdb\u884c\u66f4\u51c6\u786e\u7684\u9884\u6d4b\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6211\u4eec\u6ce8\u610f\u5230\uff0cQK\u56de\u8def\u7684\u67e5\u627e<\/span>\uff08\u4ece\u54ea\u62ff\uff09<\/span>\u548cOV\u56de\u8def\u53d6\u5c5e\u6027<\/span>\uff08\u62ff\u4ec0\u4e48\uff09<\/span>\u672c\u6765\u662f\u72ec\u7acb\u7684\u4e24\u4ef6\u4e8b\uff0c\u7406\u5e94\u53ef\u4ee5\u5206\u522b\u6307\u5b9a\u5e76\u6309\u9700\u81ea\u7531\u7ec4\u5408\uff0cMHA\u786c\u628a\u5b83\u4eec\u653e\u5230\u4e00\u4e2a\u6ce8\u610f\u529b\u5934\u7684QKOV\u91cc\u201c\u6346\u7ed1\u9500\u552e\u201d\uff0c\u9650\u5236\u4e86\u7075\u6d3b\u6027\u548c\u8868\u8fbe\u80fd\u529b\u3002\u5047\u8bbe\u6709\u4e2a\u6a21\u578b\u5b58\u5728\u6ce8\u610f\u529b\u5934A\u3001B\u3001C\u5176QK\u548cOV\u56de\u8def\u80fd\u591f\u5b8c\u6210\u4e0a\u9762\u7684\u4f8b\u5b50\uff0c\u90a3\u6362\u6210\u201c\u5c0f\u660e\u6709\u4e00\u652f\u94a2\u7b14<\/strong>…\u5c0f\u660e\u6709\u4e00\u79cd<\/strong> -> \u6587\u5177<\/strong><\/span>\u201d<\/span>\uff08head A \u5173\u6ce8\u76f8\u540c\u4e3b\u8bed\u7684\u5bbe\u8bed\u7684QK + head B \u53d6\u7c7b\u522b\u5c5e\u6027\u7684OV\uff09<\/span>\u6216\u201c\u5c0f\u7ea2\u6709\u4e00\u652f\u94a2\u7b14<\/strong>…\u5c0f\u7ea2\u80fd\u7528\u5b83<\/strong> -> \u5199\u5b57<\/strong><\/span>\u201d<\/span>\uff08head A \u5173\u6ce8\u76f8\u540c\u4e3b\u8bed\u7684\u5bbe\u8bed\u7684 QK + head C \u53d6\u7528\u9014\u5c5e\u6027\u7684OV\uff09<\/span>\uff0c\u9700\u8981\u4ea4\u53c9\u7ec4\u5408\u73b0\u6709\u6ce8\u610f\u529b\u5934\u7684QK\u548cOV\u56de\u8def\uff0c\u6a21\u578b\u5c31\u53ef\u80fd\u201c\u8f6c\u4e0d\u8fc7\u5f2f\u513f\u201d\u4e86\u3002\u8fd9\u5c31\u50cf\u662f\u5728\u80af\u5fb7\u57fa\u70b9\u9910\uff0c\u5982\u679c\u53ea\u6709\u56fa\u5b9a\u7684\u5957\u9910\u7ec4\u5408\uff0c\u800c\u987e\u5ba2\u60f3\u8981\u7684\u7ec4\u5408\u5e76\u4e0d\u5728\u83dc\u5355\u4e0a\uff0c\u90a3\u5c31\u65e0\u6cd5\u6ee1\u8db3\u9700\u6c42\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u5b9e\u9645\u4e0a\uff0c\u5373\u4f7f\u662f\u8fd9\u6837\u7b80\u5355\u7684\u4f8b\u5b50\uff0c\u6a21\u578b\u7684\u8868\u73b0\u4e5f\u5e76\u4e0d\u603b\u662f\u4ee4\u4eba\u6ee1\u610f\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u6bd4\u59826B\u4ee5\u4e0b\u7684\u6a21\u578b\uff0c\u5b83\u4eec\u7684\u8868\u73b0\u4f1a\u51fa\u4e4e\u610f\u6599\u5730\u5dee\u3002\u8fd9\u4fc3\u4f7f\u6211\u4eec\u53bb\u6df1\u5165\u5206\u6790\u539f\u56e0\uff0c\u5e76\u5bfb\u6c42\u6539\u8fdb\u7684\u65b9\u6cd5\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/section>\n

    <\/span><\/section>\n

    \n
    \n
    \n
    \n
    \n

    <\/section>\n

    <\/section>\n
    \n
    <\/svg><\/section>\n<\/section>\n<\/section>\n
    \n

    2. DCFormer\u7684\u8bbe\u8ba1\u3001\u5b9e\u73b0\u548c\u8bc4\u4f30<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

    \n

    <\/section>\n

    <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


    <\/p>\n

    \u90a3\u4e48\uff0c\u6211\u4eec\u5e94\u8be5\u5982\u4f55\u5c06QK\u548cOV\u8fdb\u884c\u52a8\u6001\u7ec4\u5408\u5462\uff1f\u6211\u4eec\u53ef\u4ee5\u5728\u591a\u5934\u6ce8\u610f\u529b\u7684\u51e0\u4e2a\u8ba1\u7b97\u6b65\u9aa4\u4e2d\u8fdb\u884c\u7ec4\u5408\u3002\u6bd4\u5982\uff0c\u5c06Wq<\/sub>\u3001Wk<\/sub>\u3001Wo<\/sub>\u3001Wv<\/sub>\u7b49\u6743\u91cd\u7ec4\u5408\uff0c\u6216\u5bf9\u8f93\u51fa\u8fdb\u884c\u7ec4\u5408\u3002\u800c\u6211\u4eec\u5728DCFomer\u4e2d\u91c7\u7528\u7684\u5176\u5b9e\u662f\u5c06\u5b83\u751f\u6210\u7684\u6ce8\u610f\u529b\u5206\u6570\u548c\u6ce8\u610f\u529b\u6743\u91cd\u77e9\u9635\u505a\u7ec4\u5408\u3002<\/span><\/section>\n

    <\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u901a\u8fc7\u4e4b\u524d\u7684\u4f8b\u5b50\u6765\u8bf4\u660e\u5728\u6ce8\u610f\u529b\u673a\u5236\u7684\u5c42\u9762\u4e0a\uff0c\u7ec4\u5408\u4e0d\u540c\u5934<\/span>\uff08Head\uff09<\/span>\u7684\u4fe1\u606f\u4f18\u52bf\u3002\u5047\u8bbe\u6211\u4eec\u6709\u4e00\u53e5\u8bdd\uff0c\u5176\u4e2d\u6bcf\u4e00\u5bf9\u6ce8\u610f\u529b\u7684\u8d77\u59cb\u8bcd\u548c\u7ec8\u70b9\u8bcd<\/span>\uff08\u4f8b\u5982\u201cpen\u201d\u548c\u201cof\u201d\uff09<\/span>\u90fd\u53ef\u4ee5\u8ba1\u7b97\u51fa\u4e00\u7ec4\u6ce8\u610f\u529b\u5206\u6570\uff0c\u5b83\u4eec\u662f\u901a\u8fc7\u6743\u91cd\u77e9\u9635Wq<\/sub>\u548cWk<\/sub>\u76f8\u4e58\u5f97\u5230\u7684\u3002\u5047\u8bbe\u6709\u516b\u4e2a\u5934<\/span>\uff08H=8\uff09<\/span>\uff0c\u90a3\u4e48\u5c31\u6709\u516b\u7ec4Wq<\/sub>\u548cWk<\/sub>\u7684\u7ec4\u5408\u3002\u8fd9\u4e9b\u7ec4\u5408\u8ba1\u7b97\u51fa\u76848\u4e2a\u6ce8\u610f\u529b\u5206\u6570\u88ab\u79f0\u4e3a\u6ce8\u610f\u529b\u5411\u91cf\uff0c\u5b83\u4eec\u51b3\u5b9a\u4e86\u4ece\u201cpen\u201d\u5230\u201cof\u201d\u5e94\u8be5\u4f20\u9012\u54ea\u4e9b\u4fe1\u606f\u3002\u4f8b\u5982\uff0c\u5047\u8bbe\u7b2c\u516d\u4e2a\u5934<\/span>\uff08Head 6\uff09<\/span>\u7684QK\u56de\u8def\u7279\u522b\u5173\u6ce8\u4e86\u201cpen\u201d\uff0c\u5728\u6b63\u5e38\u60c5\u51b5\u4e0b\uff0c\u7b2c\u516d\u4e2a\u5934\u7684\u6570\u503c\u4f1a\u8f83\u5927\uff0c\u6a21\u578b\u5c31\u4f1a\u503e\u5411\u4e8e\u4f7f\u7528\u7b2c\u516d\u4e2a\u5934\u7684OV\u56de\u8def\u6765\u4f20\u9012\u4fe1\u606f\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u4f46\u5982\u679c\u7b2c\u516d\u4e2a\u5934\u7684OV\u5e76\u4e0d\u83b7\u53d6\u6211\u4eec\u9700\u8981\u7684\u5c5e\u6027\u4fe1\u606f\uff0c\u6bd4\u5982\u201ckindOf\u201d<\/span>\uff08\u7c7b\u522b\uff09<\/span>\uff0c\u90a3\u4e48\u6211\u4eec\u5c31\u4e0d\u80fd\u7b80\u5355\u5730\u4f7f\u7528\u5b83\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u53ef\u4ee5\u8c03\u6574\u8fd9\u4e9b\u6ce8\u610f\u529b\u5206\u6570\uff0c\u5047\u8bbe\u7b2c\u56db\u4e2a\u5934<\/span>\uff08Head 4\uff09<\/span>\u7684OV\u5305\u542b\u6240\u9700\u7684\u201ckindOf\u201d\u4fe1\u606f\u3002\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5c06Head 6\u7684\u6ce8\u610f\u529b\u5206\u6570\u589e\u52a0\u7ed9Head 4\uff0c\u8ba9\u5b83\u5728\u8ba1\u7b97\u4e2d\u8d77\u5230\u66f4\u5927\u7684\u4f5c\u7528\uff0c\u4ece\u800c\u5b9e\u73b0\u5728Head 6\u7684QK\u642d\u8d77\u7684\u901a\u8def\u4e0a\u4f20\u9012Head 4\u7684OV\u83b7\u53d6\u7684\u5c5e\u6027\u4fe1\u606f\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u901a\u8fc7\u8fd9\u6837\u201c\u5077\u6881\u6362\u67f1\u201d\uff0c\u6211\u4eec\u5c31\u5b9e\u73b0\u4e86\u4e0d\u540c\u5934\u4e4b\u95f4\u7684\u4fe1\u606f\u7ec4\u5408\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u66f4\u7075\u6d3b\u5730\u5904\u7406\u4fe1\u606f\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6211\u4eec\u53ef\u4ee5\u64cd\u4f5c\u6ce8\u610f\u529b\u5411\u91cf\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u66f4\u51c6\u786e\u5730\u6355\u6349\u548c\u4f20\u9012\u5173\u952e\u4fe1\u606f\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u90a3\u4e48\uff0c\u5982\u4f55\u5c06\u4e00\u4e2a\u6ce8\u610f\u529b\u5411\u91cfA:ij<\/sub>\u8f6c\u6362\u6210\u6211\u4eec\u60f3\u8981\u7684\u53e6\u4e00\u4e2a\u5411\u91cfA:ij<\/sub>‘\u5462\uff1f\u4e00\u79cd\u7b80\u5355\u7684\u65b9\u6cd5\u662f\u901a\u8fc7\u77e9\u9635\u4e58\u6cd5\u3002\u5f53\u6211\u4eec\u5c06\u4e00\u4e2a\u5411\u91cf\u4e0e\u4e00\u4e2a\u7279\u5b9a\u7684\u77e9\u9635\u76f8\u4e58\u65f6\uff0c\u5b83\u5c31\u53ef\u4ee5\u8f6c\u6362\u6210\u53e6\u4e00\u4e2a\u5411\u91cf\u3002\u8fd9\u4e2a\u77e9\u9635\u7684\u6027\u8d28\u51b3\u5b9a\u4e86\u8f6c\u6362\u7684\u5177\u4f53\u65b9\u5f0f\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u4f8b\u5982\uff0c\u5982\u679c\u77e9\u9635\u8bbe\u8ba1\u6210\u7279\u5b9a\u7684\u5f62\u5f0f\uff0c\u5b83\u5c31\u53ef\u4ee5\u5b9e\u73b0\u7279\u5b9a\u7684\u53d8\u6362\uff0c\u6bd4\u5982\u5c06\u539f\u672c\u5c5e\u4e8e\u7b2c\u56db\u4e2a\u5934\u7684\u4fe1\u606f\u79fb\u52a8\u5230\u7b2c\u516d\u4e2a\u5934\uff0c\u6216\u8005\u53cd\u4e4b\u3002\u6b64\u5916\uff0c\u77e9\u9635\u8fd8\u53ef\u4ee5\u8c03\u6574\u5404\u4e2a\u5934\u7684\u6743\u91cd\uff0c\u6bd4\u5982\u589e\u5f3a\u6216\u51cf\u5f31\u67d0\u4e2a\u5934\u7684\u5f71\u54cd\u529b\u3002\u6bd4\u5982\uff0c\u5982\u679c\u6211\u4eec\u60f3\u8981\u51cf\u5f31\u7b2c46\u4e2a\u4f4d\u7f6e\u7684\u6743\u91cd\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u77e9\u9635\u7684\u5bf9\u89d2\u7ebf\u4e0a\u76f8\u5e94\u4f4d\u7f6e\u8bbe\u7f6e\u8f83\u5c0f\u7684\u503c\uff0c\u8fd9\u6837\u5728\u4e58\u6cd5\u64cd\u4f5c\u540e\uff0c\u8be5\u4f4d\u7f6e\u7684\u6743\u91cd\u5c31\u4f1a\u964d\u4f4e\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6211\u4eec\u5c06\u8fd9\u4e9b\u7528\u4e8e\u8c03\u6574\u548c\u7ec4\u5408\u7684\u77e9\u9635\u79f0\u4e3a\u201cComposition Map\u201d<\/span>\uff08\u7ec4\u5408\u6620\u5c04\uff09<\/span>\u3002\u8fd9\u4e9b\u6620\u5c04\u4e0d\u662f\u9488\u5bf9\u5355\u4e2a\u6ce8\u610f\u529b\u8d77\u70b9-\u7ec8\u70b9\u5bf9\uff0c\u800c\u662f\u5168\u5c40\u6027\u7684\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u5982\u679c\u8fd9\u4e9b\u7ec4\u5408\u6620\u5c04\u5bf9\u4e8e\u6240\u6709\u8d77\u70b9-\u7ec8\u70b9\u5bf9\u90fd\u662f\u76f8\u540c\u7684\uff0c\u6211\u4eec\u79f0\u4e4b\u4e3a\u201c\u9759\u6001\u201d\u7684\u7ec4\u5408\u6620\u5c04\uff0c\u5176\u6548\u679c\u7b49\u540c\u4e8e\u6269\u5c55\u4e86\u5934\u7684QK<\/span>\uff08Query-Key\uff09<\/span>\u548cOV<\/span>\uff08Output Value\uff09<\/span>\u7684\u7ef4\u5ea6\uff0c\u6bd4\u5982\u5982\u679c\u67098\u4e2a\u5934\uff0c\u5c31\u53ef\u4ee5\u5c06\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\u6269\u5c55\u5230\u539f\u6765\u76848\u500d\uff0c\u4ece\u800c\u589e\u5f3a\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u5728DCMHA\u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a\u201ccompose\u201d\u7684\u5173\u952e\u64cd\u4f5c\uff0c\u662f<\/span>\u5c06\u4e00\u4e2a\u6ce8\u610f\u529b\u5411\u91cf\u8f6c\u6362\u6210\u53e6\u4e00\u4e2a\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u5411\u91cf\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u7075\u6d3b\u7684\u6ce8\u610f\u529b\u5206\u914d\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u5177\u4f53\u6765\u8bf4\uff0ccompose\u64cd\u4f5c\u4f1a\u63a5\u6536\u4e00\u4e2a\u539f\u59cb\u7684\u6ce8\u610f\u529b\u5411\u91cf\uff0c\u5e76\u8f93\u51fa\u4e00\u4e2a\u65b0\u7684\u3001\u7ecf\u8fc7\u8f6c\u6362\u7684\u6ce8\u610f\u529b\u5411\u91cf\u3002\u5728\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u4e2d\uff0c\u6211\u4eec\u4f1a\u5728\u4e24\u4e2a\u5173\u952e\u7684\u4f4d\u7f6e\u63d2\u5165compose\u64cd\u4f5c\uff1a\u4e00\u4e2a\u5728softmax\u5f52\u4e00\u5316\u6b65\u9aa4\u4e4b\u524d\uff0c\u53e6\u4e00\u4e2a\u5728softmax\u4e4b\u540e\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0cDCMHA\u80fd\u591f\u52a8\u6001\u5730\u8c03\u6574\u6bcf\u4e2a\u5934\u7684\u8d21\u732e\uff0c\u521b\u9020\u51fa\u66f4\u52a0\u7cbe\u786e\u548c\u6709\u6548\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\u3002\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\uff0c\u800c\u4e14\u589e\u5f3a\u4e86\u5176\u5bf9\u590d\u6742\u6570\u636e\u5173\u7cfb\u7684\u6355\u6349\u80fd\u529b\uff0c\u4ece\u800c\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u90fd\u80fd\u53d6\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    Noam Shazeer\u7b49\u4eba\u7684Talking-heads attention\u5de5\u4f5c\u63d0\u51fa\u8fc7\u7c7b\u4f3c\u7684\u6ce8\u610f\u529b\u5934\u7ec4\u5408\u601d\u60f3<\/span>\uff08\u5176\u672c\u8d28\u662f\u4e0a\u6587\u7684\u9759\u6001\u7ec4\u5408\u6620\u5c04\uff09<\/span>\uff0c\u4e0e\u4ed6\u4eec\u7684\u5de5\u4f5c\u4e0d\u540c\uff0c\u6211\u4eec\u7684\u5de5\u4f5c\u5219\u66f4\u52a0\u5173\u6ce8\u52a8\u6001\u7ec4\u5408\u7684\u6982\u5ff5\u3002\u6240\u8c13\u52a8\u6001\uff0c\u610f\u5473\u7740\u6839\u636e\u4e0d\u540c\u7684\u8f93\u5165\uff0c\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u751f\u6210\u4e0d\u540c\u7684composition map\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u52a8\u6001\u8c03\u6574\u4e0d\u540c\u5934\u4e4b\u95f4\u7684\u7ec4\u5408\u65b9\u5f0f\u3002\u4f8b\u5982\uff0c\u7ed9\u5b9a\u524d\u6587\u201c\u5c0f\u660e\u6709\u53ea\u94a2\u7b14\u201d\uff0c\u5bf9\u201c\u5c0f\u660e\u6709\u4e00\u79cd -> \u6587\u5177\u201d\uff0c\u6a21\u578b\u9700\u8981\u5c06\u627e\u76f8\u540c\u4e3b\u8bed\u7684QK\u4e0e\u53d6\u5c5e\u6027\u7684OV\u7ec4\u5408\uff0c\u800c\u5bf9\u201c\u5c0f\u660e\u62ff\u5b83 -> \u5199\u5b57\u201d\uff0c\u5219\u9700\u8981\u4e0e\u53d6\u7528\u9014\u7684OV\u7ec4\u5408\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \n
    Shazeer, Noam, et al. Talking-heads attention. arXiv preprint arXiv:2003.02436<\/em> (2020).<\/span><\/section>\n<\/blockquote>\n

    <\/span><\/section>\n
    \u52a8\u6001\u7ec4\u5408\u7684\u4f18\u52bf\u5728\u4e8e\u5b83\u80fd\u591f\u66f4\u7075\u6d3b\u5730\u9002\u5e94\u4e0d\u540c\u7684\u8bed\u8a00\u6a21\u5f0f\uff0c\u4f46\u8fd9\u4e5f\u5e26\u6765\u4e86\u8ba1\u7b97\u4e0a\u7684\u6311\u6218\u3002\u56e0\u4e3a\u6a21\u578b\u9700\u8981\u4e3a\u6bcf\u4e2atoken\u751f\u6210\u4e00\u4e2a\u52a8\u6001\u7684\u7ec4\u5408\u6620\u5c04\uff0c\u8fd9\u5728\u8ba1\u7b97\u4e0a\u53ef\u80fd\u4f1a\u975e\u5e38\u6602\u8d35\u3002\u7279\u522b\u662f\u5f53\u6a21\u578b\u89c4\u6a21\u8f83\u5927\u65f6\uff0c\u9700\u8981\u5904\u7406\u7684\u53c2\u6570\u548c\u8ba1\u7b97\u91cf\u4f1a\u663e\u8457\u589e\u52a0\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5728DCMHA\u4e2d\u91c7\u7528composition map\u5206\u89e3\u7684\u65b9\u6cd5\u3002\u8fd9\u79cd\u65b9\u6cd5\u5c06compose\u64cd\u4f5c\u5206\u89e3\u4e3a\u51e0\u4e2a\u6b65\u9aa4\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u90fd\u6709\u5176\u7279\u5b9a\u7684\u4f5c\u7528\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u4f4e\u79e9\u5206\u89e3\u6765\u9650\u5236\u7ec4\u5408\u7684\u590d\u6742\u6027\uff0c\u4ece\u800c\u51cf\u5c11\u8ba1\u7b97\u5f00\u9500\u3002\u4f4e\u79e9\u610f\u5473\u7740\u6bcf\u4e2a\u7279\u5b9a\u4f4d\u7f6e\u9700\u8981\u7684\u7ec4\u5408\u6570\u91cf\u662f\u6709\u9650\u7684\uff0c\u8fd9\u6837\u53ef\u4ee5\u6709\u6548\u5730\u63a7\u5236\u8ba1\u7b97\u8d44\u6e90\u7684\u4f7f\u7528\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5f15\u5165\u4e86\u95e8\u63a7\u673a\u5236\uff0c\u5b83\u901a\u8fc7\u8c03\u6574composition map\u5bf9\u89d2\u7ebf\u4e0a\u7684\u503c\u6765\u589e\u5f3a\u6216\u51cf\u5f31\u7279\u5b9a\u5934\u7684\u5f71\u54cd\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u6a21\u578b\u5728\u4e0d\u540c\u60c5\u51b5\u4e0b\u7075\u6d3b\u8c03\u6574\u6bcf\u4e2a\u5934\u7684\u8d21\u732e\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6700\u7ec8\uff0c\u901a\u8fc7\u5c06\u8fd9\u4e9b\u4e0d\u540c\u6b65\u9aa4\u7684\u7ed3\u679c\u7ed3\u5408\u8d77\u6765\uff0c\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u4e00\u4e2a\u7efc\u5408\u7684\u8f93\u51fa\u3002\u5982\u679c\u6211\u4eec\u4ece\u6574\u4f53\u5f20\u91cf\u7684\u89d2\u5ea6\u6765\u7406\u89e3\u8fd9\u4e2a\u8fc7\u7a0b\uff0c\u5b9e\u9645\u4e0a\u662f\u5bf9\u7ec4\u5408\u5f20\u91cf\u8fdb\u884c\u4e86\u884c\u5217\u5206\u89e3\u548c\u4f4e\u79e9\u5bf9\u89d2\u5206\u89e3\uff0c\u8fd9\u4e9b\u6280\u672f\u5728\u5176\u4ed6\u9886\u57df\u4e5f\u6709\u5e94\u7528\uff0c\u4f46\u6211\u4eec\u5c06\u5176\u521b\u65b0\u6027\u5730\u5e94\u7528\u4e8e\u7ec4\u5408\u6620\u5c04\u4e0a\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u8fd9\u79cd\u65b9\u6cd5\u5f15\u5165\u4e86\u4e00\u4e9b\u989d\u5916\u7684\u53c2\u6570\uff0c\u4f46\u5176\u8ba1\u7b97\u590d\u6742\u5ea6\u76f8\u5bf9\u8f83\u5c0f\u3002\u589e\u52a0\u7684\u8ba1\u7b97\u91cf<\/span>\uff08flops\uff09<\/span>\u4e5f\u4e0d\u591a\uff0c\u5e76\u4e14\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\uff0c\u8fd9\u79cd\u590d\u6742\u5ea6\u76f8\u5bf9\u51cf\u5c11\uff0c\u5373\u6a21\u578b\u8d8a\u5927\uff0c\u589e\u52a0\u7684<\/strong>\u8ba1\u7b97<\/strong>\u590d\u6742\u5ea6\u8d8a\u5c11<\/strong>\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n
    \u63a5\u4e0b\u6765\uff0c\u8ba9\u6211\u4eec\u770b\u770b\u4e3b\u8981\u7684\u6d4b\u8bd5\u7ed3\u679c\u2014\u2014\u6027\u80fd\u7b97\u529b\u6bd4<\/strong>\u3002\u6027\u80fd\u7b97\u529b\u6bd4\u8861\u91cf\u7684\u662f\u5355\u4f4d\u7b97\u529b\u6295\u5165\u80fd\u5e26\u6765\u591a\u5c11\u6027\u80fd\u63d0\u5347\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u5728\u76f8\u540c\u7684\u8bad\u7ec3\u6570\u636e<\/span>\uff08300B tokens\uff09<\/span>\u4e0b\uff0c\u4e00\u4e2a\u6539\u8fdb\u540e\u768469\u4ebf\u53c2\u6570\u6a21\u578b\uff0c\u5176\u6548\u679c\u751a\u81f3\u8d85\u8fc7\u4e86120\u4ebf\u53c2\u6570\u7684\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u5982\u679cGPT-4o\u80fd\u591f\u5e94\u7528\u8fd9\u9879\u6280\u672f\uff0c\u63a8\u7406\u4e00\u6b21128k\u4e0a\u6587\u7684\u6210\u672c\u5c31\u53ef\u80fd\u4ece4\u5143\u964d\u4f4e\u52302\u5143\u3002\u6b64\u5916\uff0c\u5728\u4e0b\u6e38\u4efb\u52a1\u7684\u8bc4\u4f30\u4e2d\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\uff0c\u6211\u4eec\u7684\u6539\u8fdb\u6548\u679c\u66f4\u52a0\u663e\u8457\uff0c\u59826.9B\u53c2\u6570\u6a21\u578b\u4e0e2.8B\u53c2\u6570\u6a21\u578b\u7684\u6bd4\u8f83\u3002\u5728\u4fdd\u6301\u8d85\u53c2\u6570\u548c\u8bad\u7ec3\u6570\u636e\u4e00\u81f4\u7684\u60c5\u51b5\u4e0b\uff0c\u65e0\u8bba\u662f\u635f\u5931\u503c<\/span>\uff08Loss\uff09<\/span>\u8fd8\u662f\u8bed\u8a00\u6a21\u578b\u7684\u56f0\u60d1\u5ea6<\/span>\uff08PPL\uff09<\/span>\uff0cDCPythia-6.9B\u7684\u8868\u73b0\u90fd\u4f18\u4e8ePythia-6.5B\u548cPythia-12B\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n
    <\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u5c3d\u7ba1\u7406\u8bba\u4e0a\u7684\u590d\u6742\u5ea6\u5f88\u5c0f\uff0c\u4f46\u5b9e\u9645\u8bad\u7ec3\u4e2d\u6211\u4eec\u786e\u5b9e\u53d1\u73b0\u4e86\u4e00\u4e9b\u989d\u5916\u5f00\u9500\u3002\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3acompose\u64cd\u4f5c\u5bf9\u6ce8\u610f\u529b\u77e9\u9635\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u7684\u53d8\u6362\uff0c\u867d\u7136\u6bcf\u4e2a\u53d8\u6362\u7684\u8ba1\u7b97\u91cf\u4e0d\u5927\uff0c\u4f46\u7531\u4e8eAttention\u77e9\u9635\u672c\u8eab\u5f88\u5927\uff0c\u5bfc\u81f4\u4e86\u8f83\u5927\u7684I\/O\u9700\u6c42\u3002\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5728\u5b9e\u8df5\u5c42\u9762\u8fdb\u884c\u4e86\u4e00\u4e9b\u4f18\u5316\u3002\u4f18\u5316\u7ed3\u679c\u663e\u793a\uff0c\u65e0\u8bba\u662f\u5728\u8bad\u7ec3\u8fd8\u662f\u63a8\u7406\u4e2d\uff0c\u76f8\u6bd4\u4e8eTransformer\u6a21\u578b\uff0c\u8fd9\u79cd\u989d\u5916\u5f00\u9500\u662f\u53ef\u4ee5\u63a5\u53d7\u7684\u3002\u4f8b\u5982\uff0c\u572813B\u62166.9B\u6a21\u578b\u4e0a\uff0c\u5c3d\u7ba1\u63a8\u7406\u901f\u5ea6\u6162\u4e865.5%\uff0c\u4f46\u6027\u80fd\u5374\u8d85\u8fc7\u4e8613B\u6a21\u578b\u3002\u5982\u679c\u4e0e13B\u6a21\u578b\u76f8\u6bd4\uff0c\u5b9e\u9645\u4e0a\u901f\u5ea6\u662f\u53d8\u5feb\u7684\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u4ece\u8fd9\u4e2a\u89d2\u5ea6\u6765\u770b\uff0c\u968f\u7740\u7b97\u529b\u7684\u63d0\u5347\uff0c\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\u5e26\u6765\u4e86\u66f4\u5927\u7684\u6536\u76ca\uff0c\u989d\u5916\u5f00\u9500\u5b9e\u9645\u4e0a\u968f\u7740\u7b97\u529b\u7684\u589e\u52a0\u800c\u964d\u4f4e\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u6d88\u878d\u5b9e\u9a8c\uff0c\u4ee5\u8bc4\u4f30DCMHA\u4e2d\u4e94\u4e2a\u5206\u652f\u7684\u91cd\u8981\u6027\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u52a8\u6001\u7ec4\u5408\u6bd4\u9759\u6001\u7ec4\u5408\u66f4\u91cd\u8981\uff0c\u5c24\u5176\u662f\u8de8\u5934\u7684\u7ec4\u5408\uff0c\u6bd4\u5355\u4e2a\u5934\u7684\u81ea\u589e\u5f3a\u6216\u81ea\u6291\u5236<\/span>\uff08\u5373\u95e8\u63a7\uff09<\/span>\u66f4\u4e3a\u5173\u952e\uff0c\u8fd9\u7b26\u5408\u6211\u4eec\u7684\u76f4\u89c9\uff1b\u5173<\/span>\u6ce8\u4fa7\u7684\u4fe1\u606f<\/span>\uff08query-wise\uff09<\/span>\u6bd4\u88ab\u5173\u6ce8\u4fa7\u7684\u4fe1\u606f<\/span>\uff08key-wise\uff09<\/span>\u66f4\u91cd\u8981\u3002\u5373\u4f7f\u53ea\u4f7f\u7528\u4e00\u4fa7\u7684\u4fe1\u606f\uff0c\u4e5f\u80fd\u8fbe\u5230\u5f88\u597d\u7684\u6548\u679c\u3002\u57fa\u4e8e\u8fd9\u4e9b\u89c2\u5bdf\uff0c\u6211\u4eec\u53ef\u4ee5\u8fdb\u4e00\u6b65\u91c7\u7528\u4e00\u4e9b\u65b9\u6cd5\uff0c\u6bd4\u5982\u4f7f\u7528\u66f4\u591a\u7684\u5c40\u90e8\u6ce8\u610f\u529b\u5c42\uff0c\u53ea\u7528\u5173\u6ce8\u4fa7\u7684\u4fe1\u606f<\/span>\uff08QW\uff09<\/span>\uff0c\u8fd9\u4e0d\u4ec5\u80fd\u4fdd\u6301\u7c7b\u4f3c\u7684\u6548\u679c\uff0c\u8fd8\u80fd\u8fdb\u4e00\u6b65\u63d0\u9ad8\u901f\u5ea6\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u53e6\u4e00\u9879\u5b9e\u9a8c\uff0c\u6d4b\u8bd5\u4e86DCFormer\u662f\u5426\u80fd\u591f\u89e3\u51b3\u9700\u8981\u6ce8\u610f\u529b\u5934\u7ec4\u5408\u7684\u4efb\u52a1\u3002\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u6d4b\u8bd5\u96c6\uff0c\u5305\u542b\u4e86\u5404\u79cd\u7c7b\u4f3c\u4e8e\u201c\u5c0f\u660e\u6709\u652f\u94a2\u7b14\u201d\u8fd9\u6837\u7684\u6d4b\u8bd5\u7528\u4f8b\uff0c\u5e76\u8fdb\u884c\u4e86\u4e0d\u540c\u7684QK\u548cOV\u53d8\u6362\u3002\u7ed3\u679c\u8868\u660e\uff0c\u57286.9B\u53c2\u6570\u7684\u6a21\u578b\u4e0a\uff0cDCFormer\u5728\u8fd9\u79cd\u9700\u8981\u5934\u7ec4\u5408\u7684\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u786e\u5b9e\u663e\u8457\u4f18\u4e8e\u666e\u901a\u6a21\u578b\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/section>\n
    <\/section>\n

    <\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u63a5\u4e0b\u6765\uff0c\u6211\u60f3\u8c08\u8c08\u6211\u4eec\u5de5\u4f5c\u4e2d\u5b58\u5728\u7684\u4e00\u4e9b\u5c40\u9650\u6027\u3002\u9996\u5148\uff0c\u5173\u4e8e\u5b9e\u73b0\u65b9\u9762\uff0c\u6211\u4eec\u7684\u6a21\u578b\u4e3b\u8981\u662f\u4f7f\u7528JAX\u5728TPU\u4e0a\u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u867d\u7136\u6211\u4eec\u4e5f\u63d0\u4f9b\u4e86\u5728GPU\u4e0a\u4f7f\u7528JAX\u548cPyTorch\u8fdb\u884c\u8bad\u7ec3\u7684\u4ee3\u7801\uff0c\u4f46\u6211\u4eec\u5fc5\u987b\u627f\u8ba4\uff0c\u8fd9\u90e8\u5206\u4ee3\u7801\u5e76\u6ca1\u6709\u7ecf\u8fc7\u6df1\u5165\u4f18\u5316\uff0c\u56e0\u6b64\u5728GPU\u4e0a\u7684\u6027\u80fd\u53ef\u80fd\u4f1a\u5dee\u4e8eTPU\u4e0a\u7684\u6027\u80fd\uff0c\u800c\u63a8\u7406\u90e8\u5206\u662f\u5728GPU\u4e0a\u4f7f\u7528PyTorch\u6765\u6267\u884c\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    <\/section>\n

    <\/span><\/section>\n
    \u5176\u6b21\uff0c\u5173\u4e8e\u73b0\u6709\u6a21\u578b\u7684\u9002\u914d\u95ee\u9898\u3002\u6211\u4eec\u5e0c\u671b\u80fd\u591f\u5c06DCFormer\u6280\u672f\u5e94\u7528\u5230\u73b0\u6709\u7684\u6a21\u578b\u4e0a\uff0c\u800c\u4e0d\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u4e00\u4e2a\u65b0\u6a21\u578b\uff0c\u56e0\u4e3a\u4ece\u5934\u8bad\u7ec3\u7684\u5f00\u9500\u662f\u76f8\u5f53\u5927\u7684\u3002\u6211\u4eec\u5c1d\u8bd5\u8fc7\u5c06\u73b0\u6709\u7684\u6a21\u578b\uff0c\u6bd4\u5982Llama2\u6216Llama3\uff0c\u76f4\u63a5\u5e94\u7528DCFomer\u6280\u672f\uff0c\u5e76\u901a\u8fc7\u4e00\u4e9b\u5fae\u8c03\u6765\u63d0\u5347\u6027\u80fd\u3002\u4f46\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u8fd9\u79cd\u7b80\u5355\u7684\u76f4\u63a5\u5e94\u7528\u662f\u4e0d\u591f\u7684\u3002\u539f\u56e0\u5728\u4e8e\uff0c\u6211\u4eec\u7684DCFormer\u4f9d\u8d56\u4e8e\u4e0d\u540c\u5934\u4e4b\u95f4\u7684\u7ec4\u5408\uff0c\u800c\u73b0\u6709\u7684\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u5e76\u6ca1\u6709\u8003\u8651\u5230\u8fd9\u79cd\u7ec4\u5408\u5173\u7cfb\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u5934\u7684QK<\/span>\uff08Query-Key\uff09<\/span>\u9700\u8981\u4e0e\u53e6\u4e00\u4e2a\u5934\u7684OV<\/span>\uff08Output Value\uff09<\/span>\u8fdb\u884c\u7ec4\u5408\uff0c\u8fd9\u79cd\u7ec4\u5408\u7684\u524d\u63d0\u662f\u5b83\u4eec\u5fc5\u987b\u5728\u540c\u4e00\u5c42\u4e2d\u3002\u4f46\u662f\uff0c\u5728\u73b0\u6709\u7684Llama\u6a21\u578b\u4e2d\uff0c\u9700\u8981\u7ec4\u5408\u7684QK\u548cOV\u4e0d\u4e00\u5b9a\u4f4d\u4e8e\u540c\u4e00\u5c42\uff0c\u8fd9\u5c31\u5bfc\u81f4\u4e86\u65e0\u6cd5\u6709\u6548\u5b9e\u73b0\u7ec4\u5408\u3002\u8fd9\u4e0d\u4ec5\u662f\u4e00\u4e2a\u6280\u672f\u6311\u6218\uff0c\u800c\u4e14\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u73af\u5883\u4e2d\uff0c\u8de8\u5c42\u7ec4\u5408\u7684\u5b9e\u73b0\u96be\u5ea6\u548c\u5f00\u9500\u90fd\u662f\u76f8\u5f53\u5927\u7684\u3002\u56e0\u6b64\uff0c\u76ee\u524d\u6211\u4eec\u7684\u4e00\u4e2a\u5c40\u9650\u6027\u5728\u4e8e\uff0c\u867d\u7136\u8de8\u5c42\u7ec4\u5408\u5728\u6280\u672f\u4e0a\u662f\u53ef\u884c\u7684\uff0c\u4f46\u8981\u5b9e\u73b0\u5b83\u5e76\u8fbe\u5230\u826f\u597d\u7684\u6027\u80fd\uff0c\u8fd8\u5b58\u5728\u4e00\u5b9a\u7684\u96be\u5ea6\u548c\u6210\u672c\u3002<\/span><\/section>\n

    <\/span><\/section>\n

    <\/span><\/section>\n

    \n
    \n
    \n
    \n
    \n

    <\/section>\n

    <\/section>\n
    \n
    <\/svg><\/section>\n<\/section>\n<\/section>\n
    \n

    3. \u7814\u7a76\u5386\u7a0b\u548c\u672a\u6765\u5c55\u671b<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

    \n

    <\/section>\n

    <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

    <\/h5>\n
    \u7814\u7a76\u5386\u7a0b\u56de\u987e<\/span><\/strong><\/span><\/h5>\n

    <\/section>\n
    \u5173\u4e8e\u8fd9\u7bc7\u8bba\u6587\u7684\u5de5\u4f5c\uff0c\u4ee5\u4e0a\u4ecb\u7ecd\u7684\u5c31\u662f\u6211\u4eec\u7814\u7a76\u7684\u4e3b\u8981\u5185\u5bb9\u3002\u73b0\u5728\uff0c\u6211\u60f3\u7b80\u5355\u5730\u56de\u987e\u4e00\u4e0b\u6211\u4eec\u8fdb\u884c\u8fd9\u9879\u7814\u7a76\u7684\u521d\u8877\u3002\u6211\u4eec\u6700\u521d\u7684\u51fa\u53d1\u70b9\u662f\u63d0\u9ad8\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u7cfb\u5217\u4efb\u52a1\uff0c\u5e76\u5f00\u53d1\u4e86\u81ea\u5df1\u7684\u5206\u6790\u65b9\u6cd5\u548c\u5de5\u5177\u6765\u63a2\u7a76\u6a21\u578b\u7684\u5de5\u4f5c\u673a\u5236\uff1a\u6a21\u578b\u662f\u5982\u4f55\u6267\u884c\u4efb\u52a1\u7684\uff0c\u4ee5\u53ca\u5b83\u4e3a\u4f55\u80fd\u6216\u4e0d\u80fd\u5b8c\u6210\u8fd9\u4e9b\u4efb\u52a1\u3002\u6211\u4eec\u7684\u5c1d\u8bd5\u76f8\u5f53\u4e8e\u5c06\u6a21\u578b\u9006\u5411\u5de5\u7a0b\u4ece\u201c\u9ed1\u76d2\u201d\u53d8\u4e3a\u201c\u7070\u76d2\u201d\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5bf9\u5176\u8fdb\u884c\u8c03\u8bd5\u5e76\u5bfb\u627e\u6539\u8fdb\u65b9\u6848\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u8d39\u66fc\u8bf4\uff1a\u201c\u5982\u679c\u4f60\u4e0d\u80fd\u521b\u9020\u4e00\u6837\u4e1c\u897f\uff0c\u4f60\u5c31\u4e0d\u80fd\u771f\u6b63\u7406\u89e3\u5b83\u201d\uff1b\u6211\u89c9\u5f97\u53cd\u8fc7\u6765\uff1a\u5982\u679c\u4f60\u4e0d\u80fd\u7406\u89e3\u4e00\u6837\u4e1c\u897f\uff0c\u4f60\u4e5f\u65e0\u6cd5\u6539\u8fdb\u6216\u521b\u65b0\u5b83\u3002\u7136\u800c\uff0c\u8fd9\u5e76\u975e\u552f\u4e00\u7684\u9014\u5f84\u3002\u8bb8\u591a\u6770\u51fa\u7684\u7814\u7a76\u4eba\u5458\uff0c\u6bd4\u5982\u4e4b\u524d\u63d0\u5230\u7684Noam Shazeer\uff0c\u4ed6\u4eec\u7684\u8bba\u6587\u5f80\u5f80\u4e0d\u4f1a\u8be6\u7ec6\u89e3\u91ca\u8bbe\u8ba1\u7ed3\u6784\u7684\u52a8\u673a\uff0c\u6211\u4eec\u4e5f\u65e0\u6cd5\u786e\u5207\u77e5\u9053\u4ed6\u4eec\u7684\u7075\u611f\u6765\u6e90\u3002\u53ef\u80fd\u5728\u66f4\u9ad8\u5c42\u6b21\u7684\u903b\u8f91\u4e2d\uff0c\u5b58\u5728\u7740\u5176\u4ed6\u7684\u601d\u8def\u548c\u65b9\u6cd5\uff0c\u6bd4\u5982\u76f4\u89c9\u7684\u5f15\u5bfc\u3002\u5f53\u5b9e\u8df5\u7ecf\u9a8c\u548c\u89c6\u91ce\u79ef\u7d2f\u5230\u4e00\u5b9a\u7a0b\u5ea6\u65f6\uff0c\u76f4\u89c9\u53ef\u80fd\u4f1a\u81ea\u7136\u800c\u7136\u5730\u5f15\u9886\u4f60\u8d70\u5411\u6b63\u786e\u7684\u9053\u8def<\/strong>\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u7136\u800c\uff0c\u5728\u5b9e\u9645\u7814\u7a76\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u6700\u6df1\u7684\u4f53\u4f1a\u662f\u8fed\u4ee3\u901f\u5ea6\u7684\u91cd\u8981\u6027<\/strong>\u3002\u5373\u4f7f\u4f60\u7684\u60f3\u6cd5\u5728\u7406\u8bba\u4e0a\u662f\u53ef\u884c\u7684\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u5f80\u5f80\u9700\u8981\u4e0d\u65ad\u5730\u8c03\u6574\u548c\u5b9e\u9a8c\uff0c\u5feb\u901f\u5730\u5b8c\u6210\u95ed\u73af\u3002\u6211\u4eec\u603b\u5171\u8fdb\u884c\u4e86500\u591a\u6b21\u5927\u5927\u5c0f\u5c0f\u7684\u5b9e\u9a8c\uff0c\u624d\u9a8c\u8bc1\u4e86\u6211\u4eec\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5982\u679c\u4f60\u80fd\u52a0\u5feb\u8fed\u4ee3\u901f\u5ea6\uff0c\u5c06\u539f\u672c\u4e00\u5e74\u5185\u5b8c\u6210\u7684\u5b9e\u9a8c\u538b\u7f29\u5230\u534a\u5e74\uff0c\u90a3\u4e48\u4f60\u5c31\u6709\u53ef\u80fd\u5728\u534a\u5e74\u5185\u53d6\u5f97\u663e\u8457\u7684\u6210\u679c\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6b64\u5916\uff0c\u9762\u5411\u5e95\u5c42\u786c\u4ef6\u7684\u4f18\u5316\u4e5f\u662f\u81f3\u5173\u91cd\u8981\u7684<\/strong>\u3002\u5728\u6a21\u578b\u4f18\u5316\u7684\u540e\u671f\uff0c\u6211\u4eec\u53d1\u73b0\u6027\u80fd\u63d0\u5347\u7684\u74f6\u9888\u5f80\u5f80\u4e0d\u5728\u4e8ePyTorch\u6216JAX\u7b49\u6846\u67b6\u5c42\u9762\uff0c\u800c\u5728\u4e8eTPU\u3001GPU\u7b49\u5e95\u5c42\u786c\u4ef6\u3002\u6700\u7ec8\uff0c\u6211\u4eec\u53ef\u80fd\u9700\u8981\u7f16\u5199CUDA Kernel\u7b49\u5e95\u5c42\u4ee3\u7801\uff0c\u8fd9\u4f3c\u4e4e\u662f\u4e0d\u53ef\u907f\u514d\u7684\u3002\u6700\u8fd1\u8bb8\u591a\u67b6\u6784\u521b\u65b0\u7684\u5de5\u4f5c\uff0c\u5982Flash Attention\u3001Mamba\u3001GLA\u7b49\uff0c\u5b83\u4eec\u7684\u5b9e\u73b0\u90fd\u4f9d\u8d56\u4e8e\u5b9a\u5236\u5316\u7684Kernel\uff0c\u5426\u5219\u6a21\u578b\u6548\u7387\u5c06\u53d7\u5230\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u6211\u66fe\u8ddf\u56e2\u961f\u534a\u5f00\u73a9\u7b11\u8bf4\uff0c\u5f53\u524d\u67b6\u6784\u521b\u65b0\u7684\u5c3d\u5934\u662fCUDA\u7f16\u7a0b\u3002\u5982\u679c\u4f60\u7684\u60f3\u6cd5\u4ec5\u9650\u4e8ePyTorch\u3001JAX\u5c42\u9762\uff0c\u90a3\u4e48\u4f60\u7684\u53d1\u6325\u7a7a\u95f4\u53ef\u80fd\u53d7\u5230\u9650\u5236\u3002\u800c\u5982\u679c\u4f60\u638c\u63e1\u4e86CUDA\u3001Pallas\u7b49\u5e95\u5c42\u6280\u672f\uff0c\u4f60\u5c06\u6709\u66f4\u591a\u7684\u7a7a\u95f4\u53bb\u505a\u51fa\u66f4\u51fa\u8272\u7684\u5de5\u4f5c\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    2B\u9752\u5e74\u7684\u672a\u6765\u65b9\u5411<\/span><\/strong><\/span><\/h5>\n

    <\/span><\/section>\n
    \u63a5\u4e0b\u6765\uff0c\u8ba9\u6211\u4eec\u63a2\u8ba8\u4e00\u4e0b\u5e95\u5c42\u7ed3\u6784\u521b\u65b0\u7684\u53ef\u80fd\u6027\u548c\u672a\u6765\u65b9\u5411\uff0c\u6211\u5c06\u5206\u4eab\u4e00\u4e9b\u4e2a\u4eba\u770b\u6cd5\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6211\u575a\u4fe1\u5e95\u5c42\u7ed3\u6784\u7684\u521b\u65b0\u4e0d\u4ec5\u53ef\u884c\uff0c\u800c\u4e14\u662f\u4e00\u4e2a\u5145\u6ee1\u6f5c\u529b\u7684\u9886\u57df<\/strong>\u3002\u81ea\u4eceMamba\u6a21\u578b\u51fa\u73b0\u540e\uff0c\u5b83\u6fc0\u53d1\u4e86\u4e00\u7cfb\u5217\u4ee4\u4eba\u5174\u594b\u7684\u7814\u7a76\u5de5\u4f5c\uff0c\u8fd9\u4e00\u9886\u57df\u4e5f\u9010\u6e10\u53d7\u5230\u66f4\u5e7f\u6cdb\u7684\u5173\u6ce8\u3002\u5728\u8fc7\u53bb\uff0c\u8bb8\u591a\u4eba\u8ba4\u4e3aTransformer\u67b6\u6784\u5df2\u7ecf\u8fbe\u5230\u4e86\u5e95\u5c42\u7ed3\u6784\u7684\u6781\u9650\uff0c\u65e0\u9700\u8fdb\u4e00\u6b65\u6539\u8fdb\u3002\u7136\u800c\uff0cMamba\u3001DCFormer\u7b49\u5de5\u4f5c\u7684\u51fa\u73b0\u6253\u7834\u4e86\u8fd9\u4e00\u523b\u677f\u5370\u8c61\uff0c\u4e3a\u6211\u4eec\u6253\u5f00\u4e86\u65b0\u7684\u63a2\u7d22\u7a7a\u95f4\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u7279\u522b\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u6211\u4eec\u5bf9\u6df7\u5408\u67b6\u6784\u975e\u5e38\u611f\u5174\u8da3\uff0c\u5c24\u5176\u662f\u5c06Attention\u673a\u5236\u4e0eRNN\u6216\u72b6\u6001\u7a7a\u95f4\u6a21\u578b<\/span>\uff08SSM\uff09<\/span>\u7ed3\u5408\u8d77\u6765\u3002\u8fd9\u4e24\u79cd\u67b6\u6784\u5177\u6709\u4e0d\u540c\u7684\u5185\u5728\u504f\u597d<\/span>\uff08Inductive Bias\uff09<\/span>\uff0c\u5b83\u4eec\u7684\u7ed3\u5408\u80fd\u591f\u5f62\u6210\u5f3a\u5927\u7684\u6df7\u5408\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u7b80\u5355\u5730\u5c06\u4e00\u5c42Attention\u4e0e\u4e00\u5c42RNN\u7ed3\u5408\uff0c\u5c31\u80fd\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\u3002\u65e0\u8bba\u662f\u5df2\u53d1\u8868\u7684\u7814\u7a76\u8fd8\u662f\u6211\u4eec\u7684\u5b9e\u9a8c\uff0c\u90fd\u8bc1\u5b9e\u4e86\u8fd9\u79cd\u6df7\u5408\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u6b64\u5916\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0e\u4eba\u8111\u7684\u5de5\u4f5c\u65b9\u5f0f\u66f4\u4e3a\u8d34\u8fd1\u3002\u4eba\u8111\u5728\u7ef4\u6301\u5f53\u524d\u72b6\u6001\u7684\u540c\u65f6\uff0c\u80fd\u591f\u901a\u8fc7\u8fdc\u7a0bAttention\u673a\u5236\u8fdb\u884c\u8bb0\u5fc6\u68c0\u7d22\uff0c\u800c\u6df7\u5408\u67b6\u6784\u6b63\u662f\u8fd9\u79cd\u80fd\u529b\u7684\u4f53\u73b0\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6b64\u5916\uff0c\u4fe1\u606f\u6d41\u52a8\u6027\u4e5f\u662f\u6211\u4eec\u5173\u6ce8\u7684\u7126\u70b9\u3002\u65e0\u8bba\u662fMamba\u8fd8\u662fDCFormer\uff0c\u5b83\u4eec\u90fd\u901a\u8fc7\u8f93\u5165\u76f8\u5173\u7684\u65b9\u5f0f\u589e\u5f3a\u4e86\u6a21\u578b\u5185\u90e8\u7684\u4fe1\u606f\u6d41\u52a8\u3002\u5c3d\u7ba1Transformer\u6a21\u578b\u672c\u8eab\u4e5f\u80fd\u52a8\u6001\u5f62\u6210\u4fe1\u606f\u6d41\u901a\u8def\u5f84\uff0c\u4f46\u6211\u4eec\u8ba4\u4e3a\u8fd9\u4e9b\u8def\u5f84\u7684\u52a8\u6001\u6027\u8fd8\u6709\u5f85\u63d0\u9ad8\u3002\u56e0\u6b64\uff0c\u63a2\u7d22\u5982\u4f55\u4ee5\u66f4\u9ad8\u6548\u7684\u65b9\u5f0f\u4fc3\u8fdb\u4fe1\u606f\u6d41\u52a8\u901a\u8def\u7684\u52a8\u6001\u5f62\u6210\uff0c\u662f\u4e00\u4e2a\u5145\u6ee1\u5e0c\u671b\u7684\u7814\u7a76\u65b9\u5411\u3002<\/span><\/section>\n

    <\/span><\/section>\n

    <\/span><\/section>\n
    \n
    \n
    \n

    \u5b66\u8005\u7b80\u4ecb<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

    \n


    <\/p>\n<\/section>\n

    <\/section>\n
    \u8096\u8fbe<\/strong>\uff0c\u5317\u4eac\u90ae\u7535\u5927\u5b66\u8bb2\u5e08\uff0c\u5f69\u4e91\u79d1\u6280\u9996\u5e2d\u79d1\u5b66\u5bb6\uff0c\u96c6\u667a\u4ff1\u4e50\u90e8\u6838\u5fc3\u6210\u5458\u3002\u957f\u671f\u4ece\u4e8b\u6df1\u5ea6\u5b66\u4e60\u3001\u5927\u6a21\u578b\u7684\u7b97\u6cd5\u7814\u7a76\u53ca\u5176\u5728\u5929\u6c14\u9884\u62a5\u3001\u673a\u5668\u7ffb\u8bd1\u3001\u673a\u5668\u8f85\u52a9\u521b\u4f5c\u3001\u7cfb\u7edf\u548c\u4ee3\u7801\u5b89\u5168\u7b49\u9886\u57df\u7684\u5e94\u7528\u7814\u7a76\uff0c\u76f8\u5173\u5de5\u4f5c\u53d1\u8868\u5728 ICLR\u3001ICML\u3001TDSC\u7b49\u9876\u7ea7\u4f1a\u8bae\u548c\u671f\u520a\u3002\u76ee\u524d\u7814\u7a76\u5174\u8da3\u5305\u62ec\u5927\u6a21\u578b\u673a\u5236\u53ef\u89e3\u91ca\u6027\u3001\u9ad8\u6548\u6a21\u578b\u67b6\u6784\u8bbe\u8ba1\u3002<\/span><\/section>\n


    <\/span><\/p>\n

    \n
    \n
    \n

    \u56e2\u961f\u4ecb\u7ecd\u53ca\u62db\u8058\u5c97\u4f4d<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

    \n


    <\/p>\n<\/section>\n

    DCFomer\u7684\u8bba\u6587\u4f5c\u8005\u5747\u6765\u81ea\u5f69\u4e91NLP\u7b97\u6cd5\u7ec4\uff0c\u6838\u5fc3\u6210\u5458\u51fa\u81ea\u96c6\u667a\u4ff1\u4e50\u90e8\u3002\u56e2\u961f\u7ee7\u627f\u4e86\u96c6\u667a\u4ff1\u4e50\u90e8\u81ea\u7531\u8ba8\u8bba\u548c\u4e50\u4e8e\u5949\u732e\u7684\u7cbe\u795e\uff0c\u9f13\u52b1\u5f00\u653e\u3001\u771f\u8bda\u3001\u5e73\u7b49\u7684\u6c9f\u901a\uff0c\u79c9\u627f\u7740\u7cbe\u76ca\u6c42\u7cbe\u7684\u7cbe\u795e\uff0c\u81f4\u529b\u4e8e\u201c\u8ba9\u7530\u95f4\u5730\u5934\u7684\u519c\u6c11\u5927\u54e5\u201d\u548c\u201c\u5b57\u5b99\u91cc\u7684\u822a\u5929\u5458\u201d\u4e00\u6837\uff0c\u90fd\u80fd\u901a\u8fc7\u4eba\u5de5\u667a\u80fd\u4ea7\u54c1\u8ba9\u751f\u6d3b\u66f4\u7f8e\u597d\u3002<\/span><\/section>\n

    <\/span><\/section>\n
    \u6211\u4eec\u6b63\u5728\u62db\u8058\u591a\u4e2a\u7b97\u6cd5\u5c97\u4f4d\uff0c\u671f\u5f85\u4f18\u79c0\u7684\u4f60\u52a0\u5165\u6211\u4eec\u56e2\u961f\uff01<\/span><\/section>\n

    <\/span><\/section>\n
    \n
    \n
    \n
    \n
    \n
    \n

    <\/section>\n<\/section>\n
    \n
    \n
    \n
    \n
    1.\u5927\u6a21\u578b\u5bf9\u9f50&\u6d4b\u8bc4\u7b97\u6cd5\u7814\u7a76\u5458<\/span><\/strong><\/span><\/section>\n
    \u5c97\u4f4d\u804c\u8d23<\/span><\/strong><\/span><\/section>\n