{"id":36573,"date":"2025-01-14T06:00:43","date_gmt":"2025-01-13T22:00:43","guid":{"rendered":"https:\/\/17aitech.com\/?p=36573"},"modified":"2025-01-14T06:00:43","modified_gmt":"2025-01-13T22:00:43","slug":"%e6%97%a0%e9%97%ae%e8%8a%af%e7%a9%b9%e6%8f%90%e5%87%ba%e6%b7%b7%e5%90%88%e7%a8%80%e7%96%8f%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%96%b9%e6%a1%88moa%ef%bc%8c%e5%8a%a0%e9%80%9f%e9%95%bf%e6%96%87%e6%9c%ac","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36573","title":{"rendered":"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-11-08-6\" target=\"_blank\">\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347<\/a><\/p>\n<section>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u5728\u957f\u6587\u672c\u573a\u666f\u4e0b\u7684\u9700\u6c42\u4e0d\u65ad\u6d8c\u73b0\uff0c\u5176\u6838\u5fc3\u7684<strong>\u6ce8\u610f\u529b\u673a\u5236\uff08Attention Mechanism\uff09<\/strong>\u4e5f\u83b7\u5f97\u4e86\u975e\u5e38\u591a\u7684\u5173\u6ce8\u3002<\/section>\n<section><\/section>\n<section>\u6ce8\u610f\u529b\u673a\u5236\u4f1a\u8ba1\u7b97\u4e00\u5b9a\u8de8\u5ea6\u5185\u8f93\u5165\u6587\u672c\uff08\u4ee4\u724c\uff0cToken\uff09\u4e4b\u95f4\u7684\u4ea4\u4e92\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u4e0a\u4e0b\u6587\u7684\u7406\u89e3\u3002\u968f\u7740\u5e94\u7528\u7684\u53d1\u5c55\uff0c\u9ad8\u6548\u5904\u7406\u66f4\u957f\u8f93\u5165\u7684\u9700\u6c42\u4e5f\u968f\u4e4b\u589e\u957f [1][2]\uff0c\u8fd9\u5e26\u6765\u4e86\u8ba1\u7b97\u4ee3\u4ef7\u7684\u6311\u6218\uff1a\u6ce8\u610f\u529b\u9ad8\u6602\u7684\u8ba1\u7b97\u6210\u672c\u548c\u4e0d\u65ad\u589e\u957f\u7684\u952e\u503c\u7f13\u5b58\uff08KV-Cache\uff09\u4ee3\u4ef7\u3002\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u53ef\u4ee5\u6709\u6548\u7f13\u89e3\u5185\u5b58\u548c\u541e\u5410\u91cf\u7684\u6311\u6218\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u73b0\u6709\u7a00\u758f\u6ce8\u610f\u529b\u901a\u5e38\u91c7\u7528\u7edf\u4e00\u7684\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\uff0c\u5373\u5bf9\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u5934\u548c\u8f93\u5165\u957f\u5ea6\u5e94\u7528\u76f8\u540c\u7684\u7a00\u758f\u6a21\u5f0f\u3002\u8fd9\u79cd\u7edf\u4e00\u7684\u65b9\u6cd5\u96be\u4ee5\u6355\u6349\u5230\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u591a\u6837\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\uff0c\u5bfc\u81f4\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u4e0d\u540c\u7684\u7cbe\u5ea6 &#8211; \u4ee3\u4ef7\u6743\u8861\u88ab\u5ffd\u7565\u3002<\/section>\n<section><\/section>\n<section>\u6700\u8fd1\uff0c<strong>\u6765\u81ea\u6e05\u534e\u5927\u5b66\u3001\u65e0\u95ee\u82af\u7a79\u548c\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u7684\u7814\u7a76\u56e2\u961f\u53d1\u8868\u4e86\u300aMoA: Mixture of Sparse Attention for Automatic Large Language Model Compression\u300b\uff0c\u63d0\u51fa\u901a\u8fc7\u6df7\u5408\u4e0d\u540c\u7a00\u758f\u5ea6\u7684\u6ce8\u610f\u529b\u5934\uff0c\u4f7f\u7528 25% \u7684\u6ce8\u610f\u529b\u7a20\u5bc6\u5ea6\uff0c\u5c31\u53ef\u4ee5\u8bb0\u5fc6\u51e0\u4e4e 100% \u7684\u4e0a\u4e0b\u6587<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u672c\u5de5\u4f5c\u73b0\u5df2\u5f00\u6e90\uff0c\u6b22\u8fce\u4ea4\u6d41\u8ba8\u8bba\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u4ee3\u7801\uff1ahttps:\/\/github.com\/thu-nics\/MoA<\/section>\n<\/li>\n<li>\n<section>\u4e3b\u9875\uff1ahttps:\/\/nics-effalg.com\/MoA<\/section>\n<\/li>\n<li>\n<section>arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2406.14909<\/section>\n<\/li>\n<\/ul>\n<section><strong><br \/><\/strong><\/section>\n<section><strong>\u603b\u89c8<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\uff0c\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u5934\u8868\u73b0\u51fa\u5404\u5f02\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\u548c\u6269\u5c55\u89c4\u5219\uff1a\u6709\u7684\u5173\u6ce8\u5168\u5c40\u4fe1\u606f\uff0c\u6709\u7684\u5219\u805a\u7126\u5c40\u90e8\uff1b\u6709\u7684\u6ce8\u610f\u529b\u8303\u56f4\u968f\u8f93\u5165\u957f\u5ea6\u589e\u52a0\u800c\u6269\u5c55\uff0c\u6709\u7684\u5219\u4fdd\u6301\u4e0d\u53d8\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u7edf\u4e00\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u7834\u574f\u4e86\u8fd9\u4e9b\u56fa\u6709\u7684\u7279\u6027\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\uff08Mixture of Sparse Attention, MoA\uff09\u65b9\u6cd5\uff0c\u5b83\u80fd\u591f\u4e3a\u4e0d\u540c\u7684\u5934\u548c\u5c42\u5b9a\u5236\u72ec\u7279\u7684\u7a00\u758f\u6ce8\u610f\u529b\u914d\u7f6e\u3002MoA \u6784\u5efa\u4e86\u4e00\u4e2a\u5305\u542b\u591a\u79cd\u6ce8\u610f\u529b\u6a21\u5f0f\u53ca\u5176\u6269\u5c55\u89c4\u5219\u7684\u641c\u7d22\u7a7a\u95f4\u3002\u901a\u8fc7\u5206\u6790\u6a21\u578b\uff0c\u8bc4\u4f30\u6f5c\u5728\u914d\u7f6e\uff0cMoA \u53ef\u4ee5\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u627e\u5230\u6700\u4f18\u7684\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\u548c\u6269\u5c55\u89c4\u5219\u3002<\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u65e0\u9700\u4efb\u4f55\u8bad\u7ec3\uff0cMoA \u5c31\u53ef\u4ee5\u5728\u4fdd\u6301\u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6\u4e0d\u53d8\u7684\u60c5\u51b5\u4e0b\uff0c\u5c06\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u63d0\u5347\u7ea6\u00a0<strong>3.9<\/strong> \u500d\u3002\u6a21\u578b\u6548\u679c\u4e0a\uff0c\u5728 Vicuna-7B\u3001Vicuna-13B \u548c Llama3-8B \u6a21\u578b\u4e0a\uff0cMoA \u5c06\u957f\u6587\u672c\u4fe1\u606f\u68c0\u7d22\u51c6\u786e\u7387\u63d0\u9ad8\u4e86\u00a0<strong>1.5-7.1<\/strong> \u500d\uff0c\u4f18\u4e8e\u7edf\u4e00\u6ce8\u610f\u529b\u57fa\u7ebf\u65b9\u6cd5 StreamingLLM\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0cMoA \u7f29\u5c0f\u4e86\u7a00\u758f\u4e0e\u7a20\u5bc6\u6a21\u578b\u4e4b\u95f4\u7684\u80fd\u529b\u5dee\u8ddd\uff0c\u5728 50% \u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6\u4e0b\uff0c\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u57fa\u51c6\u6d4b\u8bd5\u96c6\u7684\u6700\u5927\u76f8\u5bf9\u6027\u80fd\u4e0b\u964d\u4ece\u57fa\u7ebf\u65b9\u6cd5\u7684 9%-36% \u964d\u4f4e\u81f3 5% \u4ee5\u5185\u3002<\/section>\n<section><\/section>\n<section>\u5728\u8fd0\u884c\u6548\u7387\u4e0a\uff0cMoA \u7684\u7a00\u758f\u6ce8\u610f\u529b\u4f7f\u5f97\u751f\u6210\u8fc7\u7a0b\u4e2d KV-Cache \u957f\u5ea6\u4e0d\u6269\u5927\u4fbf\u4e8e\u5185\u5b58\u7ba1\u7406\uff0c\u51cf\u5c11\u4e86\u6ce8\u610f\u529b\u8ba1\u7b97\u91cf\uff0c\u964d\u4f4e\u4e86\u5b58\u50a8\u91cf\u4ece\u800c\u53ef\u589e\u5927\u6279\u5927\u5c0f\u3002\u7ed3\u5408 CUDA GPU \u7b97\u5b50\u4f18\u5316\uff0cMoA \u76f8\u6bd4\u4e8e FlashAttention2 \u548c vLLM \u5c06 7B \u548c 13B \u7a20\u5bc6\u6a21\u578b\u7684\u751f\u6210\u541e\u5410\u91cf\u5206\u522b\u53ef\u63d0\u5347 6.6-8.2 \u4ee5\u53ca 1.7-1.9 \u500d\u3002\u65b9\u6cd5\u5728 12K \u4ee5\u5185\u7684\u8f93\u5165\u957f\u5ea6\u4e0a\u641c\u7d22\u538b\u7f29\u65b9\u6848\uff0c\u538b\u7f29\u540e\u6a21\u578b\u53ef\u4ee5\u5728\u957f\u8fbe 256K \u7684\u8f93\u5165\u957f\u5ea6\u4e0a\u9ad8\u6548\u4e14\u7cbe\u51c6\u5730\u68c0\u7d22\u4fe1\u606f\u3002<\/section>\n<section><\/section>\n<section><strong>\u80cc\u666f\u4ecb\u7ecd<\/strong><\/section>\n<section><\/section>\n<section>\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MHA\uff09\u673a\u5236\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u6838\u5fc3\u529f\u80fd\u4e4b\u4e00 [3]\u3002\u8be5\u673a\u5236\u4ece\u8f93\u5165\u5e8f\u5217\u51fa\u53d1\uff0c\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u5c06\u5176\u8f6c\u6362\u6210\u67e5\u8be2\uff08Q\uff09\u3001\u952e\uff08K\uff09\u548c\u503c\uff08V\uff09\u4e09\u4e2a\u77e9\u9635\u3002\u8fd9\u4e9b\u77e9\u9635\u4e0e\u4e4b\u524d\u5e8f\u5217\u7684\u952e\u503c\u7f13\u5b58\uff08KV-Cache\uff09\u76f8\u7ed3\u5408\uff0c\u5171\u540c\u8ba1\u7b97\u51fa\u6ce8\u610f\u529b\u77e9\u9635\uff08A\uff09\u3002\u4e3a\u4e86\u4fdd\u6301\u81ea\u56de\u5f52\u7279\u6027\uff0c\u8fd9\u4e00\u8ba1\u7b97\u8fc7\u7a0b\u4f1a\u901a\u8fc7\u56e0\u679c\u63a9\u819c\uff08M\uff09\u8fdb\u884c\u8c03\u6574\uff0c\u6700\u7ec8\u5f97\u5230\u8f93\u51fa\uff08O\uff09\u3002\u5177\u4f53\u516c\u5f0f\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-efab34b2b22b296b1a93bdc0214caedb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-efab34b2b22b296b1a93bdc0214caedb.png\"><\/a><\/section>\n<section>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u7684\u81ea\u56de\u5f52\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a\u9884\u586b\u5145\u548c\u89e3\u7801\u3002\u5728\u9884\u586b\u5145\u9636\u6bb5\uff0c\u6a21\u578b\u4f1a\u5904\u7406\u6574\u4e2a\u8f93\u5165\u5e8f\u5217\uff0c\u4ee5\u751f\u6210\u521d\u59cb\u7684\u54cd\u5e94\u4ee4\u724c\u3002\u968f\u540e\u8fdb\u5165\u89e3\u7801\u9636\u6bb5\uff0c\u6a21\u578b\u5229\u7528\u65b0\u4ea7\u751f\u7684\u4ee4\u724c\u4ee5\u53ca\u4e4b\u524d\u7f13\u5b58\u7684 K \u548c V \u77e9\u9635\uff0c\u9010\u6b65\u751f\u6210\u540e\u7eed\u4ee4\u724c\uff0c\u76f4\u81f3\u5b8c\u6210\u6574\u4e2a\u5e8f\u5217\u7684\u751f\u6210\u3002\u867d\u7136\u8fd9\u79cd\u8fed\u4ee3\u65b9\u6cd5\u6548\u679c\u663e\u8457\uff0c\u4f46\u968f\u7740 KV-Cache \u7684\u4e0d\u65ad\u6269\u5c55\uff0c\u5b83\u4e5f\u5e26\u6765\u4e86\u5185\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\u7684\u9700\u6c42\u589e\u52a0\u3002<\/section>\n<section><\/section>\n<section><strong>\u73b0\u6709\u65b9\u6cd5<\/strong><\/section>\n<section><\/section>\n<section>\u4e4b\u524d\u7684\u7814\u7a76\u901a\u8fc7\u5f15\u5165\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\u6765\u5e94\u5bf9\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u65f6\u7684\u6548\u7387\u6311\u6218\u3002\u5bf9\u4e8e\u751f\u6210\u5f0f\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u4e3b\u6d41\u7684\u7a00\u758f\u6a21\u5f0f\u662f\u91c7\u7528<strong>\u7edf\u4e00\u8de8\u5ea6\u6ed1\u7a97\uff1a\u5373\u4e0d\u8bba\u6ce8\u610f\u529b\u5934\u8fd8\u662f\u8f93\u5165\u957f\u5ea6\u5982\u4f55\uff0c\u90fd\u4f7f\u7528\u56fa\u5b9a\u3001\u5747\u5300\u8de8\u5ea6\u7684\u6ed1\u52a8\u7a97\u53e3\u63a9\u819c\uff0c\u8fd9\u6837\u6bcf\u4e2a\u6587\u672c\u4ec5\u5173\u6ce8\u5176\u90bb\u8fd1\u7684\u4e0a\u4e0b\u6587\u533a\u57df<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u8fd8\u4f1a\u5bf9\u6700\u521d\u7684\u51e0\u4e2a\u6587\u672c\u65bd\u52a0\u5168\u5c40\u6ce8\u610f\u529b\uff0c\u4ee5\u4fbf\u5b83\u4eec\u80fd\u591f\u5173\u6ce8\u5230\u6240\u6709\u5176\u4ed6\u6587\u672c\u3002\u8fd9\u79cd\u5c40\u90e8\u6ce8\u610f\u6a21\u5f0f\u901a\u8fc7\u4e22\u5f03\u5f53\u524d\u6ce8\u610f\u8de8\u5ea6\u4e4b\u5916\u7684 KV-Cache\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u957f\u5e8f\u5217\u573a\u666f\u4e0b\u7684\u5185\u5b58\u9700\u6c42 [4][5]\u3002\u539f\u5219\u4e0a\uff0c\u5c3d\u7ba1\u5355\u4e2a\u6a21\u578b\u5c42\u7684\u6ce8\u610f\u529b\u662f\u5c40\u90e8\u7684\uff0c\u4f46\u901a\u8fc7\u591a\u5c42\u6a21\u578b\u7684\u9010\u6b65\u4f20\u9012\uff0c\u6bcf\u4e2a\u8bcd\u6700\u7ec8\u90fd\u80fd\u83b7\u53d6\u5168\u5c40\u4fe1\u606f\uff0c\u4ece\u800c\u5728\u7406\u8bba\u4e0a\u53ef\u4ee5\u5b9e\u73b0\u6bd4\u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6\u66f4\u957f\u7684\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6 [6]\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u8fd9\u79cd\u7edf\u4e00\u8de8\u5ea6\u7684\u6ed1\u52a8\u7a97\u53e3\u65b9\u6cd5\u5e76\u672a\u8003\u8651\u5230\u6a21\u578b\u672c\u8eab\u7684\u7279\u6027\uff0c\u5bfc\u81f4\u5927\u578b\u6a21\u578b\u5728\u5904\u7406\u957f\u6587\u672c\u65f6\u7684\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u53d7\u5230\u9650\u5236\uff0c\u8fdb\u800c\u5f71\u54cd\u4e86\u5176\u5728\u957f\u6587\u672c\u573a\u666f\u4e0b\u7684\u8868\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u6839\u636e\u4e4b\u524d\u7684\u7814\u7a76\u5b9a\u4e49\uff0c\u672c\u5de5\u4f5c\u5c06<strong>\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6<\/strong>\u5b9a\u4e49\u4e3a\u5728\u5185\u5bb9\u68c0\u7d22\u4efb\u52a1\u4e0a\u80fd\u591f\u8fbe\u5230 90% \u4ee5\u4e0a\u7cbe\u5ea6\u7684\u6700\u5927\u8f93\u5165\u957f\u5ea6 [1][2]\u3002\u7814\u7a76\u8868\u660e\uff0c\u50cf StreamingLLM [4] \u8fd9\u6837\u7684<strong>\u7edf\u4e00\u8de8\u5ea6\u6ed1\u7a97\u65b9\u6cd5\uff0c\u5176\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u5f80\u5f80\u96be\u4ee5\u8d85\u51fa\u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6<\/strong>\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5f53\u4f7f\u7528\u8f93\u5165\u957f\u5ea6 50% \u7684\u8de8\u5ea6\u8fdb\u884c\u7a00\u758f\u6ce8\u610f\u529b\u65f6\uff0c\u7edf\u4e00\u8de8\u5ea6\u6ed1\u7a97\u65e0\u6cd5\u6709\u6548\u5730\u4ece\u7a97\u53e3\u5916\u7684\u5185\u5bb9\u4e2d\u68c0\u7d22\u4fe1\u606f\uff0c\u800c\u4e14\u8fd9\u4e00\u95ee\u9898\u968f\u7740\u8f93\u5165\u957f\u5ea6\u7684\u589e\u52a0\u800c\u6108\u53d1\u4e25\u91cd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-885d09bd8fb5e94766637db137810ac4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-885d09bd8fb5e94766637db137810ac4.png\"><\/a><\/section>\n<section><em><sup>\u5728LongEval\u6570\u636e\u96c6\u4e0a[7]\uff0c\u4f7f\u7528\u4e0d\u540c\u6ce8\u610f\u529b\u65b9\u6cd5\u7684Vicuna-7B\u6a21\u578b\u5728\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u548c\u68c0\u7d22\u4f4d\u7f6e\u7684\u68c0\u7d22\u7cbe\u5ea6\u3002\u5927\u6d77\u635e\u9488\u5b9e\u9a8c[8]\u5c06\u5927\u91cf\u7684\u952e\u503c\u5bf9\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u5728\u4e0d\u540c\u4f4d\u7f6e\u6d4b\u8bd5\u7ed9\u952e\u68c0\u7d22\u503c\u65f6\u7684\u7cbe\u5ea6\u3002(a) \u539f\u59cb\u7a20\u5bc6\u6ce8\u610f\u529b\u6a21\u578b\uff1b(b)\u7edf\u4e00\u8de8\u5ea6\u6ed1\u7a97StreamingLLM\uff0c\u6ce8\u610f\u529b\u8de8\u5ea6\u51cf\u534a\uff0c\u8d85\u51fa\u8de8\u5ea6\u65f6\u68c0\u7d22\u6548\u679c\u964d\u4f4e\uff1b(c) MoA\uff0c\u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6\u51cf\u534a\uff0c\u8d85\u51fa\u8de8\u5ea6\u65f6\u68c0\u7d22\u6548\u679c\u4f9d\u7136\u4f18\u79c0\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e0b\u56fe\u63ed\u793a\u4e86\u8fd9\u4e2a\u73b0\u8c61\u7684\u4e00\u4e2a\u53ef\u80fd\u89e3\u91ca\uff1a\u867d\u7136\u90e8\u5206\u6ce8\u610f\u529b\u5934\u4e13\u6ce8\u4e8e\u5c40\u90e8\u4e0a\u4e0b\u6587\uff0c\u4f46<strong>\u53e6\u4e00\u4e9b\u6ce8\u610f\u529b\u5934\u5219\u5173\u6ce8\u4e86\u6574\u4e2a\u8f93\u5165\u5e8f\u5217<\/strong>\u3002\u56e0\u6b64\uff0c\u91c7\u7528\u7edf\u4e00\u8de8\u5ea6\u7684\u65b9\u6cd5\u9650\u5236\u4e86\u90a3\u4e9b\u5173\u6ce8\u5168\u5c40\u4e0a\u4e0b\u6587\u7684\u5934\u7684\u6ce8\u610f\u529b\u8de8\u5ea6\uff0c\u5e76\u4e14\u7ed9\u5173\u6ce8\u5c40\u90e8\u4e0a\u4e0b\u6587\u7684\u5934\u5206\u914d\u4e86\u8fc7\u591a\u7684\u8ba1\u7b97\u548c\u5185\u5b58\u8d44\u6e90\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c<strong>\u968f\u7740\u8f93\u5165\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u67d0\u4e9b\u6ce8\u610f\u529b\u5934\u9700\u8981\u6bd4\u5176\u4ed6\u5934\u66f4\u5feb\u5730\u6269\u5c55\u5176\u6ce8\u610f\u529b\u8de8\u5ea6<\/strong>\uff0c\u4ee5\u9632\u6b62\u6027\u80fd\u663e\u8457\u4e0b\u964d\u3002\u9057\u61be\u7684\u662f\uff0c\u7edf\u4e00\u8de8\u5ea6\u7684\u65b9\u6cd5\u6ca1\u6709\u8003\u8651\u5230\u8fd9\u79cd\u5f02\u8d28\u6027\uff0c\u672a\u80fd\u9488\u5bf9\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u9700\u6c42\u5206\u522b\u8c03\u6574\u5176\u6ce8\u610f\u529b\u8303\u56f4\u3002\u53e6\u5916\uff0c\u73b0\u6709\u7684\u6a21\u578b\u538b\u7f29\u6280\u672f\u5728\u5236\u5b9a\u538b\u7f29\u7b56\u7565\u65f6\u901a\u5e38\u57fa\u4e8e<strong>\u901a\u7528\u8bed\u6599\u5e93\u548c\u4eba\u5de5\u64b0\u5199\u7684\u8f93\u51fa\u7ed3\u679c\uff0c\u8fd9\u5e76\u4e0d\u80fd\u7cbe\u786e\u5730\u53cd\u6620\u7a00\u758f\u5316\u5bf9\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u5177\u4f53\u5f71\u54cd<\/strong>\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8b35c61270e150422dfbf368da353eab.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8b35c61270e150422dfbf368da353eab.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u6765\u81eaVicuna-7B\u6a21\u578b\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u6ce8\u610f\u529b\u77e9\u9635\u793a\u4f8b\u3002\u6bcf\u4e2a\u6ce8\u610f\u529b\u77e9\u9635\u662f\u4eceLongEval\u6570\u636e\u96c6\u7684256\u4e2a\u8f93\u5165\u4e0a\u53d6\u5e73\u5747\u5f97\u5230\u7684\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u672c\u6587\u65b9\u6cd5<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u6df7\u5408\u6ce8\u610f\u529b\uff08MoA\uff09\u7684\u65b9\u6cd5\uff0c\u5b83\u662f\u4e00\u79cd\u65e0\u9700\u8bad\u7ec3\u6216\u5fae\u8c03\u7684\u5f02\u8d28\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cMoA \u5efa\u7acb\u4e86\u4e00\u5957<strong>\u5f02\u8d28\u5f39\u6027\u89c4\u5219<\/strong>\uff0c\u4f5c\u4e3a\u6ce8\u610f\u529b\u8de8\u5ea6\u7684\u641c\u7d22\u7a7a\u95f4\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\uff0cMoA \u5c06<strong>\u81ea\u52a8\u5206\u6790<\/strong>\u4e0d\u540c\u6ce8\u610f\u529b\u8de8\u5ea6\u7684\u6548\u679c\uff0c\u5e76<strong>\u6700\u4f18\u5316<\/strong>\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u8de8\u5ea6\u3002\u540c\u65f6\uff0cMoA \u7cbe\u5fc3\u8bbe\u8ba1\u4e86<strong>\u6821\u51c6\u6570\u636e\u96c6<\/strong>\uff0c\u4ee5\u786e\u4fdd\u5b83\u80fd\u7cbe\u786e\u5730\u53cd\u6620\u51fa\u7a00\u758f\u6ce8\u610f\u529b\u5bf9\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u5177\u4f53\u5f71\u54cd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-d15d8a45197417d93e551f39e7923756.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-d15d8a45197417d93e551f39e7923756.png\"><\/a><\/section>\n<section><em><sup>MoA\u7684\u6982\u89c8\uff1a(a) \u7a00\u758f\u6ce8\u610f\u529b\u7684\u641c\u7d22\u7a7a\u95f4\u6db5\u76d6\u4e86\u6ed1\u7a97\u8de8\u5ea6\u7684\u5f02\u6784\u5f39\u6027\u89c4\u5219\uff1b(b) \u81ea\u52a8\u538b\u7f29\u8fc7\u7a0b\u59cb\u4e8e\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u6821\u51c6\u6570\u636e\u96c6\u3002MoA\u901a\u8fc7\u5206\u6790\u8fd9\u4e2a\u6570\u636e\u96c6\u4e2d\u6bcf\u4e2a\u6ce8\u610f\u529b\u503c\u5bf9\u6a21\u578b\u9884\u6d4b\u7684\u5f71\u54cd\uff0c\u63ed\u793a\u4e86\u4e0d\u540c\u5019\u9009\u5f39\u6027\u89c4\u5219\u5728\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u4e0b\u7684\u51c6\u786e\u6027\u635f\u5931\u3002\u5728\u4f18\u5316\u9636\u6bb5\uff0cMoA\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u6311\u9009\u51fa\u6700\u5408\u9002\u7684\u5f39\u6027\u89c4\u5219\uff0c\u4f7f\u5f97\u5176\u5728\u6ee1\u8db3\u5e73\u5747\u8de8\u5ea6\u7ea6\u675f\u7684\u540c\u65f6\uff0c\u5c3d\u53ef\u80fd\u51cf\u5c11\u6a21\u578b\u7684\u6027\u80fd\u635f\u5931\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u5f02\u8d28\u5f39\u6027\u89c4\u5219\u7684\u641c\u7d22\u7a7a\u95f4<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u6784\u5efa MoA \u7a00\u758f\u6ce8\u610f\u529b\u6a21\u5f0f\u7684\u641c\u7d22\u7a7a\u95f4\u65f6\uff0c\u672c\u5de5\u4f5c\u5145\u5206\u8003\u8651\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u6ce8\u610f\u529b\u6a21\u5f0f\u7684\u5929\u7136\u5f02\u8d28\u6027\u548c\u5f39\u6027\u7279\u70b9\u3002\u6b63\u5982\u56fe (a) \u5c55\u793a\u7684\u90a3\u6837\uff0c\u672c\u5de5\u4f5c\u9009\u7528\u4e86\u5bf9\u786c\u4ef6\u53cb\u597d\u7684<strong>\u5f02\u8d28\u8de8\u5ea6\u6ed1\u7a97<\/strong>\u4f5c\u4e3a\u672c\u5de5\u4f5c\u7a00\u758f\u6ce8\u610f\u529b\u7684\u63a9\u819c\u3002\u540c\u4e4b\u524d\u7814\u7a76 [4][5] \u7c7b\u4f3c\uff0c\u672c\u5de5\u4f5c\u5bf9\u6700\u521d\u7684\u51e0\u4e2a\u6587\u672c\u65bd\u52a0\u4e86\u5168\u5c40\u6ce8\u610f\u529b\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u6355\u6349\u6ce8\u610f\u529b\u8de8\u5ea6\u968f\u8f93\u5165\u957f\u5ea6\u53d8\u5316\u7684\u4e0d\u540c\u8868\u73b0\uff0c\u672c\u5de5\u4f5c\u5c06\u6ce8\u610f\u529b\u5934 h \u7684\u6ce8\u610f\u529b\u8de8\u5ea6 S \u5b9a\u4e49\u4e3a\u8f93\u5165\u957f\u5ea6 N \u7684\u51fd\u6570\uff0c\u5373<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e1e0549252adcaff458c4b53003ac9be.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e1e0549252adcaff458c4b53003ac9be.png\"><\/a>\u3002\u5176\u4e2d\uff0c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6c60593ac43c993bd2efa7e4221c398b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6c60593ac43c993bd2efa7e4221c398b.png\"><\/a>\u548c\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8f87b840701da95e7d430d8eb34bb285.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8f87b840701da95e7d430d8eb34bb285.png\"><\/a>\u662f\u63a7\u5236\u57fa\u672c\u8de8\u5ea6\u548c\u5176\u968f\u8f93\u5165\u957f\u5ea6\u53d8\u5316\u7684\u6269\u5c55\u6bd4\u4f8b\u7684\u8d85\u53c2\u6570\u3002\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e791618b13954b6583f5945dd19d9a98.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e791618b13954b6583f5945dd19d9a98.png\"><\/a>\u00a0\u548c\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-b1924ce276bba52b7a62d29c244c5fd4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-b1924ce276bba52b7a62d29c244c5fd4.png\"><\/a>\u00a0\u90fd\u53ef\u4ee5\u4ece\u4e00\u7cfb\u5217\u79bb\u6563\u9009\u9879\u4e2d\u6311\u9009<\/section>\n<section><\/section>\n<section><strong>\u81ea\u52a8\u6ce8\u610f\u529b\u8de8\u5ea6\u89c4\u5219\u641c\u7d22<\/strong><\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u6709\u5f88\u591a\u5c42\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5f02\u8d28\u5f39\u6027\u89c4\u5219\u7684\u641c\u7d22\u7a7a\u95f4\u53ef\u80fd\u4f1a\u53d8\u5f97\u975e\u5e38\u5927\u3002\u56e0\u6b64\uff0c\u672c\u5de5\u4f5c\u8bbe\u8ba1\u4e86\u81ea\u52a8\u6ce8\u610f\u529b\u8de8\u5ea6\u89c4\u5219\u641c\u7d22\u65b9\u6cd5\uff0c\u6765\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6bcf\u4e00\u4e2a\u6ce8\u610f\u529b\u5934\u786e\u5b9a\u6700\u4f18\u7684\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6dfdce80dc0475e7b7da71eccec225e7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-6dfdce80dc0475e7b7da71eccec225e7.png\"><\/a>\u00a0\u548c\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9c3c62f5d0e967d6cef73739094c717a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9c3c62f5d0e967d6cef73739094c717a.png\"><\/a>\u3002\u8fd9\u5305\u62ec\u5206\u6790\u548c\u4f18\u5316\u4e24\u6b65\u3002<\/section>\n<section><\/section>\n<section><strong>\u5206\u6790<\/strong><\/section>\n<section><\/section>\n<section>\u7ed9\u5b9a\u5927\u8bed\u8a00\u6a21\u578b\uff0cMoA \u9996\u5148\u4f1a\u5728\u6821\u51c6\u6570\u636e\u96c6\u7684\u4e00\u7cfb\u5217\u6587\u672c\u4e0a\u8fdb\u884c\u5206\u6790\uff0c\u4ee5\u8bc4\u4f30\u79fb\u9664\u6bcf\u4e2a\u6ce8\u610f\u529b\u503c\u5bf9\u6a21\u578b\u6700\u7ec8\u9884\u6d4b\u7ed3\u679c\u7684\u5177\u4f53\u5f71\u54cd\u3002\u5177\u4f53\u800c\u8a00\uff0c\u672c\u5de5\u4f5c\u91c7\u7528\u4e86\u4e00\u9636\u6cf0\u52d2\u5c55\u5f00\u7684\u65b9\u6cd5\u8fdb\u884c\u5f71\u54cd\u8bc4\u4f30\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-0747e657b60b2efb32dcb07736fded06.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-0747e657b60b2efb32dcb07736fded06.png\"><\/a><\/section>\n<section>\u901a\u8fc7\u5206\u6790\u6bcf\u4e2a\u6ce8\u610f\u529b\u503c\u5bf9\u9884\u6d4b\u7ed3\u679c\u7684\u8d21\u732e\uff0c\u6211\u4eec\u53ef\u4ee5\u8ba1\u7b97\u51fa\u5728\u5f53\u524d\u8f93\u5165\u4e0b\uff0c<strong>\u4e0d\u540c\u8de8\u5ea6\u6ed1\u7a97\u5bf9\u6700\u7ec8\u9884\u6d4b\u7ed3\u679c\u7684\u5f71\u54cd<\/strong>\uff0c\u4e5f\u5c31\u662f\u8be5\u7a97\u53e3\u6240\u79fb\u9664\u7684\u6240\u6709\u6ce8\u610f\u529b\u503c\u5f71\u54cd\u7684\u603b\u548c\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u672c\u5de5\u4f5c\u63a8\u5bfc\u4e86\u6ce8\u610f\u529b\u5f71\u54cd\u7684\u5f62\u5f0f\u5316\u8868\u8fbe\uff0c\u5e76\u5229\u7528\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u53cd\u5411\u4f20\u64ad\u673a\u5236\u9ad8\u6548\u5730\u8ba1\u7b97\u6240\u9700\u7684\u504f\u5bfc\u6570\u503c\u3002\u5728\u6821\u51c6\u6570\u636e\u96c6\u7684\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u4e0a\uff0cMoA \u5bf9\u6ed1\u52a8\u7a97\u53e3\u7684\u5f71\u54cd\u5206\u522b\u53d6\u5e73\u5747\u503c\uff0c\u4ee5\u4f53\u73b0\u540c\u4e00\u79cd\u5f02\u8d28\u5f39\u6027\u89c4\u5219\u5728\u4e0d\u540c\u957f\u5ea6\u8f93\u5165\u4e0b\u7684\u5f71\u54cd\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5206\u6790\u9636\u6bb5\u5b8c\u6210\u540e\uff0cMoA \u80fd\u591f\u660e\u786e\u6bcf\u79cd\u5f02\u8d28\u5f39\u6027\u89c4\u5219\u5728\u7cbe\u5ea6\u548c\u6548\u7387\u4e4b\u95f4\u7684\u5e73\u8861\u3002\u57fa\u4e8e\u6b64\uff0cMoA \u53ef\u4ee5\u5c06\u8f83\u957f\u7684\u6ce8\u610f\u529b\u8de8\u5ea6\u5206\u914d\u7ed9\u90a3\u4e9b\u5bf9\u538b\u7f29\u66f4\u4e3a\u654f\u611f\u7684\u6ce8\u610f\u529b\u5934\uff0c\u800c\u5c06\u8f83\u77ed\u7684\u6ce8\u610f\u529b\u8de8\u5ea6\u5206\u914d\u7ed9\u90a3\u4e9b\u5bf9\u6b64\u4e0d\u592a\u654f\u611f\u7684\u6ce8\u610f\u529b\u5934\u3002<\/section>\n<section><\/section>\n<section><strong>\u4f18\u5316<\/strong><\/section>\n<section><\/section>\n<section>\u6839\u636e\u5206\u6790\u7ed3\u679c\uff0cMoA \u4f1a\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u6311\u9009\u51fa\u6700\u4f73\u5f39\u6027\u89c4\u5219\u3002\u5728\u4f18\u5316\u8fc7\u7a0b\u4e2d\uff0c\u7528\u6237\u53ef\u4ee5\u8bbe\u5b9a\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u4e0b\u7684\u6ce8\u610f\u529b\u5bc6\u5ea6\uff08\u5373\u5e73\u5747\u6ce8\u610f\u529b\u8de8\u5ea6\u4e0e\u8f93\u5165\u957f\u5ea6\u7684\u6bd4\u503c\uff09\u4f5c\u4e3a\u9650\u5236\u6761\u4ef6\uff0c\u800c MoA \u5219\u4f1a\u5728\u6ee1\u8db3\u8fd9\u4e00\u6761\u4ef6\u7684\u540c\u65f6\uff0c\u529b\u6c42\u6700\u5c0f\u5316\u5404\u957f\u5ea6\u4e0b\u7684\u9884\u6d4b\u8bef\u5dee\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9884fb0852b13fd5fac0b6702400ac19.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9884fb0852b13fd5fac0b6702400ac19.png\"><\/a><\/section>\n<section>\u5177\u4f53\u800c\u8a00\uff0cMoA \u6784\u5efa\u4e86\u4e00\u4e2a\u591a\u76ee\u6807\u4f18\u5316\u95ee\u9898\uff1a\u6700\u5c0f\u5316\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u4e0b\u7684\u9884\u6d4b\u635f\u5931\uff0c\u540c\u65f6\u786e\u4fdd\u5e73\u5747\u6ce8\u610f\u529b\u5bc6\u5ea6\u4e0d\u8d85\u8fc7\u7528\u6237\u8bbe\u5b9a\u7684\u9650\u5236\u3002MoA \u91c7\u7528\u6df7\u5408\u6574\u6570\u4f18\u5316\u65b9\u6cd5\u6c42\u89e3\u8fd9\u4e00\u591a\u76ee\u6807\u4f18\u5316\u95ee\u9898\u3002\u6c42\u89e3\u5f97\u5230\u7684\u6240\u6709\u538b\u7f29\u65b9\u6848\u5747\u8fbe\u5230\u4e86\u5e15\u7d2f\u6258\u6700\u4f18\uff0c\u610f\u5473\u7740\u5728\u6ca1\u6709\u589e\u52a0\u5176\u4ed6\u957f\u5ea6\u9884\u6d4b\u635f\u5931\u7684\u524d\u63d0\u4e0b\uff0c\u65e0\u6cd5\u8fdb\u4e00\u6b65\u964d\u4f4e\u4efb\u4f55\u7279\u5b9a\u957f\u5ea6\u7684\u9884\u6d4b\u635f\u5931\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u6709\u6548\u6cdb\u5316\u81f3\u672a\u66fe\u89c1\u8fc7\u7684\u8f93\u5165\u957f\u5ea6\uff0cMoA \u4f1a\u4ece\u6240\u6709\u5e15\u7d2f\u6258\u6700\u4f18\u538b\u7f29\u65b9\u6848\u4e2d\uff0c\u9009\u62e9\u5728\u672a\u89c1\u8fc7\u7684\u9a8c\u8bc1\u957f\u5ea6\u4e0a\u635f\u5931\u6700\u5c0f\u7684\u65b9\u6848\u4f5c\u4e3a\u6700\u7ec8\u91c7\u7528\u7684\u538b\u7f29\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u901a\u8fc7\u81ea\u52a8\u6ce8\u610f\u529b\u8de8\u5ea6\u89c4\u5219\u641c\u7d22\uff0cMoA \u5728\u9075\u5b88\u7528\u6237\u5b9a\u4e49\u7684\u5bc6\u5ea6\u7ea6\u675f\u7684\u540c\u65f6\uff0c\u627e\u5230\u5408\u9002\u7684\u5f02\u8d28\u5f39\u6027\u89c4\u5219\u6765\u6700\u5c0f\u5316\u7531\u6ce8\u610f\u529b\u7a00\u758f\u5316\u5f15\u8d77\u7684\u51c6\u786e\u6027\u635f\u5931\u3002<\/section>\n<section><\/section>\n<section><strong>\u6821\u51c6\u6570\u636e\u96c6\u7684\u8bbe\u8ba1\u4e0e\u9009\u62e9<\/strong><\/section>\n<section><\/section>\n<section>\u540c\u65f6\uff0cMoA \u4e5f\u5f3a\u8c03\u4e86\u6570\u636e\u5de5\u7a0b\u5728\u5927\u8bed\u8a00\u6a21\u578b\u538b\u7f29\u4e2d\u7684\u91cd\u8981\u6027\u3002\u672c\u5de5\u4f5c\u53d1\u73b0\uff0c\u4f7f\u7528\u5177\u6709\u957f\u8ddd\u79bb\u4f9d\u8d56\u6027\u7684\u6570\u636e\u96c6\u5e76\u53c2\u8003\u539f\u59cb\u5927\u8bed\u8a00\u6a21\u578b\u7684\u54cd\u5e94\u5bf9\u4e8e\u51c6\u786e\u5206\u6790\u538b\u7f29\u7684\u5f71\u54cd\u81f3\u5173\u91cd\u8981\u3002<\/section>\n<section><\/section>\n<section>\u672c\u5de5\u4f5c\u6307\u51fa\u4e86\u5e38\u7528\u7684\u901a\u7528\u8bed\u8a00\u5efa\u6a21\u6570\u636e\u96c6\u7684\u4e3b\u8981\u95ee\u9898\u3002\u8fd9\u7c7b\u6570\u636e\u96c6\uff0c\u4f8b\u5982\u4eba\u7c7b\u7f16\u5199\u7684\u6587\u672c\u8bed\u6599\u5e93\uff0c\u901a\u8fc7\u5728\u6574\u4e2a\u8bed\u6599\u5e93\u4e0a\u8fdb\u884c\u4e0b\u4e00\u4e2a\u8bcd\u9884\u6d4b\u4f5c\u4e3a\u76d1\u7763\u4fe1\u53f7\u3002\u4f46\u662f\u8fd9\u4e3b\u8981\u6355\u83b7\u7684\u662f\u4e34\u8fd1\u4e0a\u4e0b\u6587\u4e4b\u95f4\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\uff0c\u800c\u5ffd\u7565\u4e86\u957f\u671f\u4e0a\u4e0b\u6587\u4f9d\u8d56\u6027\uff0c\u65e0\u6cd5\u89e3\u51b3\u50cf\u957f\u8ddd\u79bb\u68c0\u7d22\u8fd9\u6837\u7684\u5168\u5c40\u6ce8\u610f\u529b\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u540c\u65f6\uff0c\u6a21\u578b\u54cd\u5e94\u548c\u4eba\u7c7b\u7f16\u5199\u7684\u76d1\u7763\u4e4b\u95f4\u5b58\u5728\u663e\u8457\u7684\u4e0d\u5bf9\u9f50\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u540c\u4e00\u4e2a\u95ee\u9898\uff0c\u4eba\u7c7b\u53ef\u80fd\u4f1a\u56de\u7b54 &#8216;Blue&#8217;\uff0c\u800c\u6a21\u578b\u53ef\u80fd\u4f1a\u56de\u7b54 &#8216;The blue color&#8217;\u3002\u4f7f\u7528\u4eba\u7c7b\u7684\u7b54\u6848\u8fdb\u884c\u76d1\u7763\uff0c\u6ce8\u610f\u529b\u5f71\u54cd\u662f\u57fa\u4e8e\u9884\u6d4b &#8216;Blue&#8217; \u7684\u6982\u7387\u8f6c\u79fb\u91cf\u5316\u7684\uff0c\u8fd9\u4e0e\u6700\u7ec8\u76ee\u6807\u80cc\u9053\u800c\u9a70\uff0c\u5373\u96be\u4ee5\u4fdd\u6301\u539f\u59cb\u6a21\u578b\u9884\u6d4b &#8216;The&#8217; \u7684\u5173\u952e\u6ce8\u610f\u529b\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u672c\u5de5\u4f5c\u6784\u5efa\u957f\u8ddd\u79bb\u4f9d\u8d56\u5e76\u901a\u8fc7\u4e0e\u539f\u59cb\u6a21\u578b\u5bf9\u9f50\u6765\u589e\u5f3a\u6821\u51c6\u6570\u636e\u96c6\u3002\u901a\u8fc7\u4e0b\u8868\u53ef\u4ee5\u53d1\u73b0\uff0c\u8fd9\u79cd\u6570\u636e\u96c6\u6784\u5efa\u65b9\u5f0f\u53ef\u4ee5\u51c6\u786e\u53cd\u6620\u6ce8\u610f\u529b\u5f71\u54cd\uff0c\u663e\u8457\u63d0\u9ad8\u538b\u7f29\u540e\u7684\u6a21\u578b\u7684\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-239f53f6dd0c2f8973c2947b6b49fe69.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-239f53f6dd0c2f8973c2947b6b49fe69.png\"><\/a><\/section>\n<section><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/section>\n<section><\/section>\n<section><strong>\u7cbe\u5ea6<\/strong><\/section>\n<section><\/section>\n<section>MoA \u5b9e\u9a8c\u5728\u591a\u79cd\u6a21\u578b\uff08Vicuna-{7B, 13B) \u548c Llama-3-{8B, 70B}\uff09\u548c\u591a\u79cd\u57fa\u51c6\u6d4b\u8bd5\uff08\u957f\u4e0a\u4e0b\u6587\u68c0\u7d22\uff0c\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\uff09\u4e0a\u548c\u4e4b\u524d\u7684\u9759\u6001\u548c\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\uff08StreamingLLM [4],H2O [9] \u548c InfLLM [12]\uff09\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-f24bbfd3bcd6aafe37d1e32dd6b79694.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-f24bbfd3bcd6aafe37d1e32dd6b79694.png\"><\/a><\/section>\n<section>MoA \u8d85\u8d8a\u57fa\u7ebf\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\uff0c\u5e76\u5728 50% \u7684\u5e73\u5747\u6ce8\u610f\u529b\u5bc6\u5ea6\u4e0b\u8fbe\u5230\u4e86\u4e0e\u539f\u59cb\u7a20\u5bc6\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u6211\u4eec\u8ba1\u7b97\u4e86\u7a00\u758f\u6a21\u578b\u76f8\u5bf9\u4e8e\u539f\u59cb\u6a21\u578b\u7684\u6027\u80fd\u4e0b\u964d\u3002<\/section>\n<section><\/section>\n<section><strong>\u5bf9\u4e8e\u957f\u4e0a\u4e0b\u6587\u68c0\u7d22\u68c0\u7d22\u4efb\u52a1<\/strong>\uff0cMoA \u6700\u5927\u76f8\u5bf9\u51c6\u786e\u6027\u4e0b\u964d\u4e3a 8%\uff0c\u8fdc\u5c0f\u4e8e StreamingLLM\u3001InfLLM \u548c \u00a0H2O \u7684 87%\u300158% \u548c 44%\u3002\u5e73\u5747\u6765\u8bf4\uff0cMoA \u7684\u76f8\u5bf9\u51c6\u786e\u6027\u4e0b\u964d\u5728 1% \u4ee5\u4e0b\uff0c\u800c StreamingLLM\u3001InfLLM \u548c H2O \u7684 \u00a051%\u300141% \u548c 20%\u3002<\/section>\n<section><\/section>\n<section>\u5982\u4e0b\u56fe (a) \u6240\u793a\uff0cMoA \u5c06\u5176\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u5230\u6ce8\u610f\u529b\u8de8\u5ea6\u7684\u5927\u7ea6 3.9 \u500d\u3002\u56fe (b) \u663e\u793a\uff0c\u5728\u56fa\u5b9a\u7684 8k \u8f93\u5165\u957f\u5ea6\u4e0b\uff0cMoA \u53ea\u9700\u8981 25% \u6ce8\u610f\u529b\u5c31\u53ef\u4ee5\u8fbe\u5230 90% \u4ee5\u4e0a\u7684\u68c0\u7d22\u51c6\u786e\u6027\u3002\u56fe (c) \u663e\u793a MoA \u5728 12k \u957f\u5ea6\u5185\u538b\u7f29\u540e\uff0c\u53ef\u4ee5\u4fdd\u6301\u548c\u539f\u7a20\u5bc6\u6a21\u578b\u4e00\u81f4\u7684\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8af1e1fea3481bf5535df850d23ea1b4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8af1e1fea3481bf5535df850d23ea1b4.png\"><\/a><\/section>\n<section><em><sup>\u5728LongEval\u4e0a\u8fdb\u884c\u4e0a\u4e0b\u6587\u68c0\u7d22\u51c6\u786e\u6027\u6d4b\u8bd5\u3002\u4f7f\u7528Vicuna-7B\u6a21\u578b\uff1a<\/sup><\/em><sup><em>(a)\u6539\u53d8\u6ce8\u610f\u529b\u8de8\u5ea6\uff0c\u6bd4\u8f83\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c<\/em><em>(b)\u5c06\u8f93\u5165\u957f\u5ea6\u8bbe\u5b9a\u4e3a8k\uff0c\u6bd4\u8f83\u4e0d\u540c\u6ce8\u610f\u529b\u5bc6\u5ea6\u4e0b\u7684\u68c0\u7d22\u51c6\u786e\u6027\u3002<\/em><\/sup><em><sup>\u4f7f\u7528Llama3-8B\u6a21\u578b\uff1a(c)\u5c06\u5bc6\u5ea6\u8bbe\u5b9a\u4e3a50%\uff0c\u6bd4\u8f83\u8f93\u5165\u957f\u5ea6\u589e\u52a0\u65f6\u7684\u68c0\u7d22\u51c6\u786e\u6027\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u5bf9\u4e8e\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u4efb\u52a1<\/strong>\uff0c\u5728\u4f7f\u7528 LV-Eval \u548c LongBench \u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cMoA \u53ea\u663e\u793a\u4e86\u6700\u5927 5% \u548c 3% \u7684\u76f8\u5bf9\u5206\u6570\u4e0b\u964d\uff0c\u800c StreamingLLM \u5219\u5206\u522b\u6700\u5927\u4e0b\u964d\u4e86 36% \u548c 27%\uff1bInfLLM \u6700\u5927\u4e0b\u964d\u4e86 17% \u548c 5%\uff1bH2O \u6700\u5927\u4e0b\u964d\u4e86 9% \u548c 4%<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e411b2c29e83022ff13f288cc7843c51.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-e411b2c29e83022ff13f288cc7843c51.png\"><\/a><\/section>\n<section><em><sup>\u4e0d\u540c\u6ce8\u610f\u529b\u65b9\u6cd5\u5728 50% \u5bc6\u5ea6\u4e0b\u8fdb\u884c (a) LV-Eval\u548c(b) LongBench \u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u57fa\u51c6\u6d4b\u8bd5\u3002\u6d4b\u8bd5\u4f7f\u7528 Vicuna-7B\u548c13B\u6a21\u578b\uff0c\u4ee5\u53ca Llama3-70B \u6a21\u578b\u3002\u5206\u6570\u76f8\u5bf9\u4e8e\u539f\u59cb\u7a20\u5bc6\u6a21\u578b\u8fdb\u884c\u5f52\u4e00\u5316\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u957f\u4e0a\u4e0b\u6587\u6cdb\u5316<\/strong>\u3002\u901a\u8fc7\u5728 12k \u957f\u5ea6\u5185\u8fdb\u884c\u538b\u7f29\uff0cMoA \u80fd\u591f\u6709\u6548\u6cdb\u5316\u5230 32k-256k \u7684\u957f\u5ea6\u3002\u5982\u4e0b\u8868\u6240\u793a\uff0c\u5728\u8303\u5316\u7684\u957f\u5ea6\u4e0a\uff0cMoA \u7684\u68c0\u7d22\u7cbe\u5ea6\u6bd4 InfLLM \u548c StreamingLLM \u9ad8 1.9-3.3 \u500d\uff0cLV-Eval \u8bc4\u5206\u9ad8 1.2-1.4 \u500d\uff0c\u5c55\u73b0\u51fa\u4e0e\u539f\u59cb\u7a20\u5bc6\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u5982\u4e0b\u8868\u6240\u793a\uff0c<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9207dca6717b5640c2c768764c6ffadb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-9207dca6717b5640c2c768764c6ffadb.png\"><\/a><\/section>\n<section>\u6d88\u878d\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u8bc4\u4f30\u4e86\u4e0d\u540c\u7a0b\u5ea6\u7684\u7a00\u758f\u6ce8\u610f\u529b\u6df7\u5408\u5bf9\u4e8e\u6700\u7ec8\u6027\u80fd\u7684\u5f71\u54cd\u3002\u4ece\u57fa\u672c\u7684\u7edf\u4e00\u8de8\u5ea6\u6ed1\u7a97\u5f00\u59cb\uff0c\u901a\u8fc7\u4f9d\u6b21\u5f15\u5165\u4e0d\u540c\u7a0b\u5ea6\u7684\u5f02\u8d28\u6027\uff08\u5c42\u95f4\uff0c\u6ce8\u610f\u529b\u5934\u95f4\uff0c\u8f93\u5165\u957f\u5ea6\u95f4\uff09\uff0c\u6700\u7ec8\u6a21\u578b\u6027\u80fd\u5728\u4e0d\u65ad\u63d0\u5347\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8a341d031897154857828687ef7ff7f0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-8a341d031897154857828687ef7ff7f0.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u5728 25% \u6ce8\u610f\u529b\u5bc6\u5ea6\u4e0b\u5bf9\u641c\u7d22\u7a7a\u95f4\u8fdb\u884c\u6d88\u878d\u5b9e\u9a8c\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u6548\u7387<\/strong><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u4f7f\u7528 Huggingface \u6846\u67b6\u652f\u6301\u7684 FlashAttention [10] \u548c vLLM \u6846\u67b6\u4f5c\u4e3a\u57fa\u7ebf\uff0c\u5bf9\u6bd4 MoA \u7684\u6548\u7387\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u8868\u6bd4\u8f83\u4e86 MoA \u76f8\u5bf9\u4e8e\u5404\u79cd\u6ce8\u610f\u529b\u673a\u5236\u548c LLM \u6846\u67b6\u7684\u8fd0\u884c\u6548\u7387\uff0c\u5e76\u5bf9 MoA \u7684\u6bcf\u4e2a\u8bbe\u8ba1\u5e26\u6765\u7684\u6548\u7387\u63d0\u5347\u8fdb\u884c\u4e86\u6d88\u878d\u5206\u6790\u3002\u5728 50% \u6ce8\u610f\u529b\u5bc6\u5ea6\u4e0b\uff0cMoA \u76f8\u6bd4 FlashAttention2 \u5c06\u89e3\u7801\u541e\u5410\u91cf\u63d0\u5347\u4e86 6.6-8.2 \u500d\u3002\u76f8\u6bd4 H2O \u548c InfLLM\uff0c\u89e3\u7801\u541e\u5410\u91cf\u63d0\u5347\u4e86 1.2-4.0 \u500d\u3002\u4e0e\u5305\u542b\u9ad8\u5ea6\u7cfb\u7edf\u7ea7\u4f18\u5316\u7684 vLLM \u6846\u67b6 [11] \u76f8\u6bd4\uff0cMoA \u4ecd\u5b9e\u73b0\u4e86 1.7-1.9 \u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\u3002MoA \u8fd8\u5c06 GPU \u603b\u5185\u5b58\u51cf\u5c11 1.2 \u5230 1.4 \u500d\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e9b\u541e\u5410\u91cf\u7684\u63d0\u5347\u6765\u6e90\u4e8e\u56db\u4e2a\u4e3b\u8981\u56e0\u7d20\uff1a\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u9759\u6001 KV-Cache\uff08\u7ea6 3.0 \u500d\uff09\uff1b\u7531\u4e8e\u7a00\u758f\u6027\u51cf\u5c11\u4e86\u6ce8\u610f\u529b\u8ba1\u7b97\uff08\u7ea6 1.5 \u500d\uff09\uff1b\u8f83\u5c0f\u7684 KV-Cache \u5185\u5b58\u652f\u6301\u4e86\u66f4\u5927\u7684\u6279\u5927\u5c0f\uff08\u7ea6 1.4 \u500d\uff09\uff1b\u4ee5\u53ca\u6211\u4eec\u9488\u5bf9 MoA \u5f02\u6784\u6ce8\u610f\u529b\u6240\u5b9e\u73b0\u7684 CUDA GPU \u7b97\u5b50\u4f18\u5316\uff08\u7ea6 1.2 \u500d\uff09\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-03b258056ddae47d2266fdbfd681db38.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-03b258056ddae47d2266fdbfd681db38.png\"><\/a><\/p>\n<section><em><sup>\u4e0d\u540c\u6846\u67b6\u57287B\u548c13B\u6a21\u578b\u4e0a\u7684\u6548\u7387\u5206\u6790\u3002MoA \u6bcf\u4e2a\u8bbe\u8ba1\u5e26\u6765\u7684\u6548\u7387\u63d0\u5347\u901a\u8fc7\u6d88\u878d\u5206\u6790\u5206\u4e3a\u56db\u4e2a\u90e8\u5206\u3002\u6240\u6709\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\u90fd\u4f7f\u752850%\u7684\u6ce8\u610f\u529b\u5bc6\u5ea6\u3002\u89e3\u7801\u541e\u5410\u91cf\u5728A100-80GB GPU \u663e\u5b58\u80fd\u5bb9\u7eb3\u7684\u6700\u5927\u6279\u5927\u5c0f\u4e0b\u8fdb\u884c\u8bc4\u4f30\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u4f5c\u8005\u4ecb\u7ecd<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u8bba\u6587\u7684\u5171\u540c\u4e00\u4f5c\u662f\u6e05\u534e\u5927\u5b66\u7535\u5b50\u5de5\u7a0b\u7cfb NICS-EFC \u5b9e\u9a8c\u5ba4\u7684\u5085\u5929\u4e88\u3001\u9ec4\u6d69\u5cf0\u548c\u5b81\u96ea\u5983\uff0c\u4ed6\u4eec\u6765\u81ea NICS-EFC \u5b9e\u9a8c\u5ba4\u7684 EffAlg \u56e2\u961f\u548c\u65e0\u95ee\u82af\u7a79\uff08Infinigence AI\uff09\u3002NICS-EFC \u5b9e\u9a8c\u5ba4\u7531\u6c6a\u7389\u6559\u6388\u5e26\u9886\uff0c\u5b9e\u9a8c\u5ba4\u7684\u9ad8\u6548\u7b97\u6cd5\u56e2\u961f\uff08Efficient Algorithm Team\uff0cEffAlg\uff09\u7531\u5b81\u96ea\u5983\u52a9\u7406\u7814\u7a76\u5458\u5e26\u9886\u3002EffAlg \u56e2\u961f\u7684\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\u9ad8\u6548\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\uff0c\u56e2\u961f\u7f51\u7ad9\u4e3a https:\/\/nics-effalg.com\/<\/section>\n<section><\/section>\n<section><strong>\u5f15\u7528<\/strong><\/section>\n<section><\/section>\n<section><sup>[1] Chen, Shouyuan, et al. &#8220;Extending Context Window of Large Language Models via Positional Interpolation.&#8221; ArXiv, 2023, abs\/2306.15595, https:\/\/api.semanticscholar.org\/CorpusID:259262376.<\/sup><\/section>\n<section><sup>[2] Tworkowski, Szymon, et al. &#8220;Focused Transformer: Contrastive Training for Context Scaling.&#8221; ArXiv, 2023, abs\/2307.03170, https:\/\/api.semanticscholar.org\/CorpusID:259360592.<\/sup><\/section>\n<section><sup>[3] Vaswani, Ashish, et al. &#8220;Attention is all you need.&#8221; Advances in Neural Information Processing Systems, vol. 30, 2017.<\/sup><\/section>\n<section><sup>[4] Xiao, Guangxuan, et al. &#8220;Efficient Streaming Language Models with Attention Sinks.&#8221; The Twelfth International Conference on Learning Representations, 2024.<\/sup><\/section>\n<section><sup>[5] Han, Chi, et al. &#8220;Lm-infinite: Simple on-the-fly length generalization for large language models.&#8221; arXiv preprint arXiv:2308.16137, 2023.<\/sup><\/section>\n<section><sup>[6] Zaheer, Manzil, et al. &#8220;Big bird: Transformers for longer sequences.&#8221; Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 17283-17297.<\/sup><\/section>\n<section><sup>[7] Li, Dacheng, et al. &#8220;How Long Can Open-Source LLMs Truly Promise on Context Length?&#8221; lmsys.org, June 2023, https:\/\/lmsys.org\/blog\/2023-06-29-longchat.<\/sup><\/section>\n<section><sup>[8] Fu, Yao, et al. &#8220;Data Engineering for Scaling Language Models to 128K Context.&#8221; ArXiv, 2024, abs\/2402.10171, https:\/\/api.semanticscholar.org\/CorpusID:267682361.<\/sup><\/section>\n<section><sup>[9] Zhang, Zhenyu (Allen), et al. &#8220;H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models.&#8221; ArXiv, 2023, abs\/2306.14048, https:\/\/api.semanticscholar.org\/CorpusID:259263947.<\/sup><\/section>\n<section><sup>[10] Dao, Tri, et al. &#8220;FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness.&#8221; Advances in Neural Information Processing Systems, 2022.<\/sup><\/section>\n<section><sup>[11] Kwon, Woosuk, et al. &#8220;Efficient Memory Management for Large Language Model Serving with PagedAttention.&#8221; Proceedings of the 29th Symposium on Operating Systems Principles, 2023, https:\/\/api.semanticscholar.org\/CorpusID:261697361.<\/sup><\/section>\n<section><sup>[12] Xiao, Chaojun et al. \u201cInfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory.\u201d (2024).<\/sup><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-11-08-6\" target=\"_blank\">\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8 [&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":[65],"class_list":["post-36573","post","type-post","status-publish","format-standard","hentry","category-news","tag-65"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347 - \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=36573\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=36573\",\"url\":\"https:\/\/17aitech.com\/?p=36573\",\"name\":\"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=36573#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=36573#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png\",\"datePublished\":\"2025-01-13T22:00:43+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=36573#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=36573\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=36573#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=36573#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347\"}]},{\"@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":"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347 - \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=36573","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36573","url":"https:\/\/17aitech.com\/?p=36573","name":"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36573#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36573#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png","datePublished":"2025-01-13T22:00:43+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36573#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36573"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36573#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/11\/frc-ac52c88b5b309a8c0a9f0d4a5a8445d5.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36573#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u65e0\u95ee\u82af\u7a79\u63d0\u51fa\u6df7\u5408\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848MoA\uff0c\u52a0\u901f\u957f\u6587\u672c\u751f\u6210\uff0c\u5b9e\u73b0\u6700\u9ad88\u500d\u541e\u5410\u7387\u63d0\u5347"}]},{"@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\/36573","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=36573"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36573\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36573"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36573"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36573"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}