{"id":36299,"date":"2025-01-09T03:00:13","date_gmt":"2025-01-08T19:00:13","guid":{"rendered":"https:\/\/17aitech.com\/?p=36299"},"modified":"2025-01-09T03:00:13","modified_gmt":"2025-01-08T19:00:13","slug":"mit%e9%9f%a9%e6%9d%be%e5%9b%a2%e9%98%9f%e9%95%bf%e4%b8%8a%e4%b8%8b%e6%96%87llm%e6%8e%a8%e7%90%86%e9%ab%98%e6%95%88%e6%a1%86%e6%9e%b6duoattention%ef%bc%9a%e5%8d%95gpu%e5%ae%9e%e7%8e%b0330%e4%b8%87token","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36299","title":{"rendered":"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-24-3\" target=\"_blank\">MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<p><strong>\u672c\u6587\u7b2c\u4e00\u4f5c\u8005\u8096\u5149\u70dc\u662f\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u7535\u5b50\u5de5\u7a0b\u4e0e\u8ba1\u7b97\u673a\u79d1\u5b66\u7cfb\uff08MIT EECS\uff09\u7684\u4e09\u5e74\u7ea7\u535a\u58eb\u751f\uff0c\u5e08\u4ece\u97e9\u677e\u6559\u6388\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u6df1\u5ea6\u5b66\u4e60\u52a0\u901f\uff0c\u5c24\u5176\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u52a0\u901f\u7b97\u6cd5\u8bbe\u8ba1\u3002\u4ed6\u5728\u6e05\u534e\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u4e0e\u6280\u672f\u7cfb\u83b7\u5f97\u672c\u79d1\u5b66\u4f4d\u3002\u4ed6\u7684\u7814\u7a76\u5de5\u4f5c\u5e7f\u53d7\u5173\u6ce8\uff0cGitHub\u4e0a\u7684\u9879\u76ee\u7d2f\u8ba1\u83b7\u5f97\u8d85\u8fc79000\u9897\u661f\uff0c\u5e76\u5bf9\u4e1a\u754c\u4ea7\u751f\u4e86\u91cd\u8981\u5f71\u54cd\u3002\u4ed6\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ecSmoothQuant\u548cStreamingLLM\uff0c\u8fd9\u4e9b\u6280\u672f\u548c\u7406\u5ff5\u5df2\u88ab\u5e7f\u6cdb\u5e94\u7528\uff0c\u96c6\u6210\u5230NVIDIA TensorRT-LLM\u3001HuggingFace\u53caIntel Neural Compressor\u7b49\u5e73\u53f0\u4e2d\u3002\u672c\u6587\u7684\u6307\u5bfc\u8001\u5e08\u4e3a\u97e9\u677e\u6559\u6388\uff08https:\/\/songhan.mit.edu\/\uff09<\/strong><strong><br \/><\/strong><\/p>\n<section>TL;DR\uff1aDuoAttention \u901a\u8fc7\u5c06\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6ce8\u610f\u529b\u5934\u5206\u4e3a\u68c0\u7d22\u5934\uff08Retrieval Heads\uff0c\u9700\u8981\u5b8c\u6574 KV \u7f13\u5b58\uff09\u548c\u6d41\u5f0f\u5934\uff08Streaming Heads\uff0c\u53ea\u9700\u56fa\u5b9a\u91cf KV \u7f13\u5b58\uff09\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u7684\u6548\u7387\uff0c\u663e\u8457\u51cf\u5c11\u5185\u5b58\u6d88\u8017\u3001\u540c\u65f6\u63d0\u9ad8\u89e3\u7801\uff08Decoding\uff09\u548c\u9884\u586b\u5145\uff08Pre-filling\uff09\u901f\u5ea6\uff0c\u540c\u65f6\u5728\u957f\u77ed\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u4fdd\u6301\u4e86\u51c6\u786e\u7387\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-07776e81a4466b72e2ebdccf72b463f7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-07776e81a4466b72e2ebdccf72b463f7.png\"><\/a><\/p>\n<\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2410.10819<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u4e3b\u9875\u53ca\u4ee3\u7801\uff1ahttps:\/\/github.com\/mit-han-lab\/duo-attention<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5355 GPU \u5b9e\u73b0 330 \u4e07 Token \u4e0a\u4e0b\u6587\u63a8\u7406\u6f14\u793a\u89c6\u9891\uff1a<a href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650940101&amp;idx=1&amp;sn=1356fe395915221ccfa70a7f71580af1&amp;chksm=84e7e6bbb3906fad6bbfa8ff845194adfd204e97967c7914d2e41b7e4c460d7643ed60075f4f&amp;token=1079059104&amp;lang=zh_CN#rd\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-1415eaa8fe2da12b9ecaa6d47c5594d9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-1415eaa8fe2da12b9ecaa6d47c5594d9.png\"><\/a><\/a><\/section>\n<section><span data-cover=\"http%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FKmXPKA19gWicEuwW1cDLUhOZLVy5cVFHia98TGtEXuC0MVicaicticqHm0JvWWFJ6rCvo0kMp6LscB1lQXoGjE1T0AA%2F0%3Fwx_fmt%3Djpeg\" data-mpvid=\"wxv_3692223443798982659\" data-ratio=\"1.7391304347826086\" data-src=\"https:\/\/mp.weixin.qq.com\/mp\/readtemplate?t=pages\/video_player_tmpl&amp;auto=0&amp;vid=wxv_3692223443798982659\" data-vh=\"371.8125\" data-vidtype=\"2\" data-vw=\"661\" data-w=\"640\"><\/p>\n<div data-key=\"wxv_3692223443798982659\">\n<div data-v-fd7c1e60=\"\">\n<div data-v-f5e35cd2=\"\" data-v-fd7c1e60=\"\"><\/div>\n<\/div>\n<\/div>\n<p><\/span><\/section>\n<section>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Models\uff0cLLMs\uff09\u5728\u5404\u7c7b\u4efb\u52a1\u4e2d\u7684\u5e7f\u6cdb\u5e94\u7528\uff0c\u5c24\u5176\u662f\u5728\u957f\u4e0a\u4e0b\u6587\uff08Long-Context\uff09\u573a\u666f\u4e2d\u5904\u7406\u6d77\u91cf\u6587\u672c\u4fe1\u606f\uff0c\u5982\u4f55\u5728\u4fdd\u8bc1\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\u51cf\u5c11\u5185\u5b58\u548c\u8ba1\u7b97\u6210\u672c\uff0c\u6210\u4e3a\u4e86\u4e00\u4e2a\u4e9f\u5f85\u89e3\u51b3\u7684\u96be\u9898\u3002\u4e3a\u6b64\uff0c\u6765\u81ea MIT\u3001\u6e05\u534e\u5927\u5b66\u3001\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u3001\u7231\u4e01\u5821\u5927\u5b66\u548c NVIDIA \u7684\u7814\u7a76\u56e2\u961f\u8054\u5408\u63d0\u51fa\u4e86 <strong>DuoAttention \u6846\u67b6<\/strong>\u3002\u8fd9\u9879\u521b\u65b0\u6280\u672f\u901a\u8fc7\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6ce8\u610f\u529b\u673a\u5236\uff08Attention Mechanism\uff09\u8fdb\u884c\u7cbe\u7ec6\u5316\u8bbe\u8ba1\uff0c\u6781\u5927\u63d0\u9ad8\u4e86\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u7684\u6548\u7387\uff0c\u5e76\u5927\u5e45\u964d\u4f4e\u4e86\u5185\u5b58\u9700\u6c42\uff0c\u5728\u4e0d\u727a\u7272\u6a21\u578b\u51c6\u786e\u6027\u7684\u524d\u63d0\u4e0b\uff0c\u63a8\u52a8\u4e86 LLM \u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u7684\u53d1\u5c55\u3002<\/section>\n<section>\u00a0<\/section>\n<section><strong>\u7814\u7a76\u80cc\u666f\uff1a\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u7684\u6311\u6218<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff08\u5982 Llama\u3001GPT \u7b49\uff09\u5728\u591a\u8f6e\u5bf9\u8bdd\u3001\u957f\u6587\u6863\u6458\u8981\u3001\u89c6\u9891\u548c\u89c6\u89c9\u4fe1\u606f\u7406\u89e3\u7b49\u4efb\u52a1\u4e2d\u9700\u8981\u5904\u7406\u5927\u91cf\u5386\u53f2\u4fe1\u606f\uff0c\u8fd9\u4e9b\u4efb\u52a1\u5f80\u5f80\u6d89\u53ca\u6570\u5341\u4e07\u751a\u81f3\u4e0a\u767e\u4e07\u4e2a token \u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u4f8b\u5982\uff0c\u5904\u7406\u4e00\u7bc7\u5c0f\u8bf4\u3001\u6cd5\u5f8b\u6587\u6863\u6216\u89c6\u9891\u8f6c\u5f55\u5185\u5bb9\uff0c\u53ef\u80fd\u9700\u8981\u5206\u6790\u767e\u4e07\u7ea7\u522b\u7684 token\u3002\u7136\u800c\uff0c\u4f20\u7edf\u7684<strong>\u5168\u6ce8\u610f\u529b\u673a\u5236\uff08Full Attention\uff09<\/strong>\u8981\u6c42\u6a21\u578b\u4e2d\u7684\u6bcf\u4e2a token \u90fd\u8981\u5173\u6ce8\u5e8f\u5217\u4e2d\u7684\u6240\u6709\u524d\u5e8f token\uff0c\u8fd9\u5bfc\u81f4\u4e86\u89e3\u7801\u65f6\u95f4\u7ebf\u6027\u589e\u52a0\uff0c\u9884\u586b\u5145\uff08Pre-Filling\uff09\u65f6\u95f4\u5448\u4e8c\u6b21\u589e\u957f\uff0c\u540c\u65f6\uff0cKV \u7f13\u5b58\uff08Key-Value Cache\uff09\u7684\u5185\u5b58\u6d88\u8017\u4e5f\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u6210\u7ebf\u6027\u589e\u957f\u3002\u5f53\u4e0a\u4e0b\u6587\u8fbe\u5230\u6570\u767e\u4e07 token \u65f6\uff0c\u6a21\u578b\u7684\u8ba1\u7b97\u8d1f\u62c5\u548c\u5185\u5b58\u6d88\u8017\u5c06\u8fbe\u5230\u96be\u4ee5\u627f\u53d7\u7684\u5730\u6b65\u3002<\/section>\n<section><\/section>\n<section><strong>DuoAttention \u7684\u521b\u65b0\u8bbe\u8ba1<\/strong><\/section>\n<section><\/section>\n<section>\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0cDuoAttention \u6846\u67b6\u63d0\u51fa\u4e86\u521b\u65b0\u6027\u7684 \u201c\u68c0\u7d22\u5934\uff08Retrieval Heads\uff09\u201d \u4e0e \u201c\u6d41\u5f0f\u5934\uff08Streaming Heads\uff09\u201d \u7684\u5206\u79bb\u65b9\u6cd5\u3002\u8fd9\u4e00\u8bbe\u8ba1\u7684\u6838\u5fc3\u7406\u5ff5\u662f\uff1a\u5e76\u975e\u6240\u6709\u7684\u6ce8\u610f\u529b\u5934\uff08Attention Heads\uff09\u5728\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u65f6\u90fd\u9700\u8981\u4fdd\u7559\u5b8c\u6574\u7684 KV \u7f13\u5b58\u3002\u7814\u7a76\u56e2\u961f\u901a\u8fc7\u5927\u91cf\u5b9e\u9a8c\u53d1\u73b0\uff0c\u5728\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u4efb\u52a1\u4e2d\uff0c\u53ea\u6709\u4e00\u5c0f\u90e8\u5206\u6ce8\u610f\u529b\u5934\uff0c\u5373 \u201c\u68c0\u7d22\u5934\u201d\uff0c\u9700\u8981\u5bf9\u5168\u90e8 token \u8fdb\u884c\u5173\u6ce8\uff0c\u4ee5\u83b7\u53d6\u4e0a\u4e0b\u6587\u4e2d\u7684\u5173\u952e\u4fe1\u606f\u3002\u800c\u5927\u591a\u6570\u6ce8\u610f\u529b\u5934\uff0c\u5373 \u201c\u6d41\u5f0f\u5934\u201d\uff0c\u53ea\u9700\u5173\u6ce8\u6700\u8fd1\u7684 token \u548c\u6ce8\u610f\u529b\u6c47\u70b9\uff08Attention Sinks\uff09\uff0c\u4e0d\u9700\u8981\u5b58\u50a8\u5168\u90e8\u7684\u5386\u53f2 KV \u72b6\u6001\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-4695a9d6b420afde20a3daa3c5b6a9ec.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-4695a9d6b420afde20a3daa3c5b6a9ec.png\"><\/a><\/p>\n<\/section>\n<section>\u56fe 1 \u5c55\u793a\u4e86\u5728 Llama-2-7B \u6a21\u578b\u4e0a\u4f7f\u7528\u5168\u6ce8\u610f\u529b\u673a\u5236\u7684\u6ce8\u610f\u529b\u56fe\uff08Attention Maps\uff09\u3002\u4ece\u56fe\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c<strong>\u68c0\u7d22\u5934\uff08Retrieval Heads\uff09\u6355\u83b7\u4e86\u4e0a\u4e0b\u6587\u4e2d\u5982 &#8220;best&#8221;\u3001&#8221;fruit&#8221; \u548c &#8220;orange&#8221; \u7b49\u5173\u952e\u4fe1\u606f\uff0c\u8fd9\u4e9b\u4fe1\u606f\u5bf9\u4e8e\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u800c\u9700\u8981\u5b8c\u6574\u7684 KV \u7f13\u5b58\u3002<\/strong>\u800c\u6d41\u5f0f\u5934\uff08Streaming Heads\uff09\u5219\u4e3b\u8981\u5173\u6ce8\u6700\u8fd1\u7684 token \u548c\u6ce8\u610f\u529b\u6c47\u70b9\uff0c\u4e0d\u9700\u8981\u4fdd\u7559\u6240\u6709\u5386\u53f2\u4fe1\u606f\u3002<\/section>\n<section><\/section>\n<section><strong>DuoAttention \u7684\u5de5\u4f5c\u539f\u7406<\/strong><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-b31f103b688bf3763f3e344e033fcd4d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-b31f103b688bf3763f3e344e033fcd4d.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 2 \u8bf4\u660e\u4e86 DuoAttention \u7684\u57fa\u672c\u5de5\u4f5c\u539f\u7406\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6846\u67b6\u901a\u8fc7\u4ee5\u4e0b\u51e0\u79cd\u5173\u952e\u673a\u5236\u6765\u4f18\u5316\u63a8\u7406\u8fc7\u7a0b\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u68c0\u7d22\u5934\u7684 KV \u7f13\u5b58\u4f18\u5316<\/strong>\uff1aDuoAttention \u4e3a\u68c0\u7d22\u5934\u4fdd\u7559\u5b8c\u6574\u7684 KV \u7f13\u5b58\uff0c\u8fd9\u4e9b\u5934\u5bf9\u957f\u8ddd\u79bb\u4f9d\u8d56\u4fe1\u606f\u7684\u6355\u6349\u81f3\u5173\u91cd\u8981\u3002\u5982\u679c\u5bf9\u8fd9\u4e9b\u5934\u7684 KV \u7f13\u5b58\u8fdb\u884c\u526a\u88c1\uff0c\u5c06\u5bfc\u81f4\u6a21\u578b\u6027\u80fd\u4e25\u91cd\u4e0b\u964d\u3002\u56e0\u6b64\uff0c\u68c0\u7d22\u5934\u9700\u8981\u5bf9\u4e0a\u4e0b\u6587\u4e2d\u7684\u6240\u6709 token \u4fdd\u6301 \u201c\u5168\u6ce8\u610f\u529b\uff08Full Attention\uff09\u201d\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u6d41\u5f0f\u5934\u7684\u8f7b\u91cf\u5316 KV \u7f13\u5b58<\/strong>\uff1a\u6d41\u5f0f\u5934\u5219\u4e3b\u8981\u5173\u6ce8\u6700\u8fd1\u7684 token \u548c\u6ce8\u610f\u529b\u6c47\u70b9\u3002\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u53ea\u9700\u8981\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684 KV \u7f13\u5b58\uff08Constant-Length KV Cache\uff09\uff0c\u4ece\u800c\u51cf\u5c11\u4e86 KV \u7f13\u5b58\u5bf9\u5185\u5b58\u7684\u9700\u6c42\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0cDuoAttention \u80fd\u591f\u4ee5\u8f83\u4f4e\u7684\u8ba1\u7b97\u548c\u5185\u5b58\u4ee3\u4ef7\u5904\u7406\u957f\u5e8f\u5217\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/section>\n<\/li>\n<\/ul>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-67d6984d9da1aebb14a4b0fcd8527ed5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-67d6984d9da1aebb14a4b0fcd8527ed5.png\"><\/a><\/p>\n<\/section>\n<ul>\n<li>\n<section><strong>\u68c0\u7d22\u5934\u7684\u81ea\u52a8\u8bc6\u522b<\/strong>\uff1a\u4e3a\u4e86\u51c6\u786e\u533a\u5206\u54ea\u4e9b\u5934\u662f\u68c0\u7d22\u5934\uff0cDuoAttention \u63d0\u51fa\u4e86\u4e00\u79cd\u8f7b\u91cf\u5316\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u4f7f\u7528\u5408\u6210\u6570\u636e\u96c6\u6765\u8bad\u7ec3\u6a21\u578b\u81ea\u52a8\u8bc6\u522b\u91cd\u8981\u7684\u68c0\u7d22\u5934\u3002\u8fd9\u79cd\u4f18\u5316\u7b56\u7565\u901a\u8fc7\u5bc6\u7801\u53ec\u56de\u4efb\u52a1\uff08Passkey Retrieval\uff09\uff0c\u786e\u5b9a\u54ea\u4e9b\u6ce8\u610f\u529b\u5934\u5728\u4fdd\u7559\u6216\u4e22\u5f03 KV \u7f13\u5b58\u540e\u5bf9\u6a21\u578b\u8f93\u51fa\u6709\u663e\u8457\u5f71\u54cd\u3002\u6700\u7ec8\uff0cDuoAttention \u5728\u63a8\u7406\u65f6\u6839\u636e\u8fd9\u4e00\u8bc6\u522b\u7ed3\u679c\uff0c\u4e3a\u68c0\u7d22\u5934\u548c\u6d41\u5f0f\u5934\u5206\u522b\u5206\u914d\u4e0d\u540c\u7684 KV \u7f13\u5b58\u7b56\u7565\u3002<\/section>\n<\/li>\n<\/ul>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8575540074457fe02b827b19edc2533d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8575540074457fe02b827b19edc2533d.png\"><\/a><\/p>\n<\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 3 \u5c55\u793a\u4e86 DuoAttention \u4f7f\u7528\u7684\u5408\u6210\u6570\u636e\u96c6\u4e2d\u7684\u4e00\u4e2a\u6837\u4f8b\u3002\u56fe 4 \u5c55\u793a\u4e86 DuoAttention \u6700\u7ec8\u786e\u5b9a LLM \u4e2d\u5404\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u7c7b\u522b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u6027\u80fd\u4e0e\u51c6\u786e\u7387\u5b9e\u9a8c<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>\u4e3a\u4e86\u9a8c\u8bc1 DuoAttention \u6846\u67b6\u7684\u6709\u6548\u6027\uff0c\u7814\u7a76\u56e2\u961f\u5728\u591a\u79cd\u4e3b\u6d41 LLM \u67b6\u6784\u4e0a\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u5b9e\u9a8c\u8bc4\u4f30\uff0c\u5305\u62ec Llama-2\u3001Llama-3 \u548c Mistral \u6a21\u578b\u3002\u5b9e\u9a8c\u4e0d\u4ec5\u6d4b\u8bd5\u4e86 DuoAttention \u5728\u5185\u5b58\u4e0e\u8ba1\u7b97\u6548\u7387\u4e0a\u7684\u63d0\u5347\uff0c\u8fd8\u901a\u8fc7\u957f\u4e0a\u4e0b\u6587\u548c\u77ed\u4e0a\u4e0b\u6587\u4efb\u52a1\u5bf9\u6a21\u578b\u7684\u51c6\u786e\u7387\u8fdb\u884c\u4e86\u5168\u9762\u6d4b\u8bd5\u3002<\/section>\n<section><\/section>\n<section><strong>1.\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u8bc4\u4f30<\/strong>\uff1a\u5728 Needle-in-a-Haystack\uff08NIAH\uff09\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cDuoAttention \u5728\u6781\u6df1\u7684\u4e0a\u4e0b\u6587\u6761\u4ef6\u4e0b\u8868\u73b0\u5353\u8d8a\uff0c\u4fdd\u6301\u4e86\u9ad8\u7cbe\u5ea6\uff0c\u5e76\u5728\u5904\u7406 1048K \u4e2a token \u7684\u957f\u4e0a\u4e0b\u6587\u65f6\uff0c\u4f9d\u7136\u80fd\u591f\u4fdd\u6301\u7a33\u5b9a\u7684\u51c6\u786e\u7387\uff0c\u800c\u5176\u4ed6\u65b9\u6cd5\u7531\u4e8e\u4e22\u5931\u5173\u952e\u4fe1\u606f\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u663e\u8457\u3002\u5728 14 \u4e2a LongBench \u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cDuoAttention \u5c55\u73b0\u4e86\u5728\u4e0d\u540c\u4efb\u52a1\u4e0b\u7684\u5f3a\u5927\u6cdb\u5316\u80fd\u529b\uff0c\u80fd\u591f\u4ee5\u8f83\u4f4e\u7684 KV \u7f13\u5b58\u9884\u7b97\uff0c\u63d0\u4f9b\u63a5\u8fd1\u5168\u6ce8\u610f\u529b\u673a\u5236\u7684\u51c6\u786e\u6027\u3002\u5728\u591a\u5934\u6ce8\u610f\u529b\u6a21\u578b\uff08MHA\uff09\u4e0a\uff0cDuoAttention \u4f7f\u7528 25% \u7684 KV \u7f13\u5b58\u9884\u7b97\u5373\u53ef\u5728\u591a\u6570\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e0e\u5168\u7f13\u5b58\u76f8\u5f53\u7684\u6548\u679c\uff0c\u800c\u5728\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u6a21\u578b\uff08GQA\uff09\u4e0a\uff0c50% \u7684 KV \u7f13\u5b58\u9884\u7b97\u5373\u53ef\u7ef4\u6301\u9ad8\u7cbe\u5ea6\u8868\u73b0\u3002<\/section>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d162e3cfacb77364cf0f9134e5577395.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d162e3cfacb77364cf0f9134e5577395.png\"><\/a><\/p>\n<\/section>\n<section><strong>2.\u77ed\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u8bc4\u4f30<\/strong>\uff1a\u5728 MMLU\uff08\u591a\u9879\u9009\u62e9\u9898\uff09\u3001MBPP\uff08\u7f16\u7a0b\u80fd\u529b\uff09\u548c MT-Bench\uff08\u5e2e\u52a9\u80fd\u529b\uff09\u7b49\u77ed\u4e0a\u4e0b\u6587\u57fa\u51c6\u4e0a\uff0cDuoAttention \u4e5f\u8868\u73b0\u51fa\u8272\u3002\u5728\u4f7f\u7528 50% \u6d41\u5f0f\u5934\u7684\u60c5\u51b5\u4e0b\uff0cDuoAttention \u7684\u8868\u73b0\u51e0\u4e4e\u4e0e\u5168\u6ce8\u610f\u529b\u673a\u5236\u4e00\u81f4\uff0c\u4fdd\u6301\u4e86 LLM \u5728\u77ed\u6587\u672c\u4efb\u52a1\u4e0a\u7684\u539f\u59cb\u80fd\u529b\u3002\u4f8b\u5982\uff0c\u5728 MMLU \u57fa\u51c6\u4e0a\uff0cDuoAttention \u4ec5\u4ee5 0.03% \u7684\u5dee\u8ddd\uff0879.35% \u5bf9\u6bd4 79.38%\uff09\u5b9e\u73b0\u4e86\u4e0e\u5168\u6ce8\u610f\u529b\u673a\u5236\u7684\u76f8\u8fd1\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-a320e24669e9613b4f1d221c0056c406.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-a320e24669e9613b4f1d221c0056c406.png\"><\/a><\/section>\n<section><strong>\u5185\u5b58\u4e0e\u6548\u7387\u7684\u63d0\u5347<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<ul>\n<li>\n<section><strong>\u5185\u5b58\u6d88\u8017\u663e\u8457\u964d\u4f4e<\/strong>\uff1aDuoAttention \u5728\u591a\u5934\u6ce8\u610f\u529b\u6a21\u578b\uff08Multi-Head Attention\uff0cMHA\uff09\u4e0a\u5c06\u5185\u5b58\u6d88\u8017\u51cf\u5c11\u4e86\u00a0<strong>2.55 \u500d<\/strong>\uff0c\u5728\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u6a21\u578b\uff08Grouped-Query Attention\uff0cGQA\uff09\u4e0a\u51cf\u5c11\u4e86\u00a0<strong>1.67 \u500d<\/strong>\u3002\u8fd9\u662f\u7531\u4e8e\u5bf9\u6d41\u5f0f\u5934\u91c7\u7528\u4e86\u8f7b\u91cf\u5316\u7684 KV \u7f13\u5b58\u7b56\u7565\uff0c\u4f7f\u5f97\u5373\u4f7f\u5728\u5904\u7406\u767e\u4e07\u7ea7\u522b\u7684\u4e0a\u4e0b\u6587\u65f6\uff0c\u6a21\u578b\u7684\u5185\u5b58\u5360\u7528\u4f9d\u7136\u4fdd\u6301\u5728\u8f83\u4f4e\u6c34\u5e73\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u89e3\u7801\uff08Decoding\uff09\u548c\u9884\u586b\u5145\uff08Pre-Filling\uff09\u901f\u5ea6\u63d0\u5347<\/strong>\uff1aDuoAttention \u7684\u89e3\u7801\u901f\u5ea6\u5728 MHA \u6a21\u578b\u4e2d\u63d0\u5347\u4e86<strong>\u00a02.18 \u500d<\/strong>\uff0c\u5728 GQA \u6a21\u578b\u4e2d\u63d0\u5347\u4e86 <strong>1.50 \u500d<\/strong>\u3002\u5728\u9884\u586b\u5145\u65b9\u9762\uff0cMHA \u548c GQA \u6a21\u578b\u7684\u901f\u5ea6\u5206\u522b\u52a0\u5feb\u4e86 <strong>1.73 \u500d<\/strong>\u548c<strong>\u00a01.63 \u500d<\/strong>\uff0c\u6709\u6548\u51cf\u5c11\u4e86\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u4e2d\u7684\u9884\u586b\u5145\u65f6\u95f4\u3002<\/section>\n<\/li>\n<\/ul>\n<section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-50ef5abc484d4d325caa88e516a7afff.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-50ef5abc484d4d325caa88e516a7afff.png\"><\/a><\/p>\n<\/section>\n<ul>\n<li>\n<section><strong>\u767e\u4e07\u7ea7 token \u5904\u7406\u80fd\u529b<\/strong>\uff1a\u7ed3\u5408<strong>\u00a04 \u6bd4\u7279\u91cf\u5316\uff08Quantization\uff09\u6280\u672f<\/strong>\uff0c DuoAttention \u5b9e\u73b0 Llama-3-8B \u5728\u5355\u4e2a A100 GPU \u4e0a\u5904\u7406\u9ad8\u8fbe <strong>330 \u4e07 token\u00a0<\/strong>\u7684\u4e0a\u4e0b\u6587\uff0c\u8fd9\u4e00\u7ed3\u679c\u662f\u6807\u51c6\u5168\u6ce8\u610f\u529b\u673a\u5236\u7684 6.4 \u500d\u3002\u00a0<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-2a7c610508bb411bc80ee5c5b9dea01f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-2a7c610508bb411bc80ee5c5b9dea01f.png\"><\/a><\/section>\n<section><strong>\u5e94\u7528\u573a\u666f\u4e0e\u672a\u6765\u5c55\u671b<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>DuoAttention \u6846\u67b6\u4e3a\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u7684\u5e94\u7528\u573a\u666f\u5e26\u6765\u4e86\u5de8\u5927\u7684\u53d8\u9769\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u5927\u89c4\u6a21\u4e0a\u4e0b\u6587\u5904\u7406\u7684\u4efb\u52a1\u4e2d\u8868\u73b0\u7a81\u51fa\uff0c\u5305\u62ec\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u591a\u8f6e\u5bf9\u8bdd\u7cfb\u7edf\uff08Multi-Turn Dialogues\uff09<\/strong>\uff1aDuoAttention \u4f7f\u5bf9\u8bdd\u6a21\u578b\u80fd\u591f\u9ad8\u6548\u5904\u7406\u957f\u65f6\u95f4\u5bf9\u8bdd\u8bb0\u5f55\uff0c\u4ece\u800c\u66f4\u597d\u5730\u7406\u89e3\u7528\u6237\u4e0a\u4e0b\u6587\uff0c\u63d0\u5347\u4ea4\u4e92\u4f53\u9a8c\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u957f\u6587\u6863\u5904\u7406\u4e0e\u6458\u8981\u751f\u6210<\/strong>\uff1a\u5728\u6587\u6863\u5206\u6790\u3001\u6cd5\u5f8b\u6587\u672c\u5904\u7406\u3001\u4e66\u7c4d\u6458\u8981\u7b49\u4efb\u52a1\u4e2d\uff0cDuoAttention \u6781\u5927\u51cf\u5c11\u5185\u5b58\u5360\u7528\uff0c\u540c\u65f6\u4fdd\u6301\u9ad8\u7cbe\u5ea6\uff0c\u4f7f\u957f\u6587\u6863\u5904\u7406\u66f4\u52a0\u53ef\u884c\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u89c6\u89c9\u4e0e\u89c6\u9891\u7406\u89e3<\/strong>\uff1a\u5728\u6d89\u53ca\u5927\u91cf\u5e27\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u5904\u7406\u7684\u89c6\u89c9\u548c\u89c6\u9891\u4efb\u52a1\u4e2d\uff0cDuoAttention \u4e3a\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08Visual Language Models\uff0cVLMs\uff09\u63d0\u4f9b\u4e86\u9ad8\u6548\u63a8\u7406\u65b9\u6848\uff0c\u663e\u8457\u63d0\u5347\u4e86\u5904\u7406\u901f\u5ea6\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f\u671f\u671b DuoAttention \u6846\u67b6\u80fd\u591f\u7ee7\u7eed\u63a8\u52a8 LLM \u5728\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u9886\u57df\u7684\u53d1\u5c55\uff0c\u5e76\u4e3a\u66f4\u591a\u5b9e\u9645\u5e94\u7528\u573a\u666f\u5e26\u6765\u663e\u8457\u63d0\u5347\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-24-3\" target=\"_blank\">MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587 [&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-36299","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>MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406 - \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=36299\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=36299\",\"url\":\"https:\/\/17aitech.com\/?p=36299\",\"name\":\"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=36299#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=36299#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png\",\"datePublished\":\"2025-01-08T19:00:13+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=36299#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=36299\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=36299#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=36299#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406\"}]},{\"@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":"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406 - \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=36299","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36299","url":"https:\/\/17aitech.com\/?p=36299","name":"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36299#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36299#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png","datePublished":"2025-01-08T19:00:13+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36299#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36299"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36299#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-414caac692e0e09975f3dbaf53bf7e1b.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36299#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"MIT\u97e9\u677e\u56e2\u961f\u957f\u4e0a\u4e0b\u6587LLM\u63a8\u7406\u9ad8\u6548\u6846\u67b6DuoAttention\uff1a\u5355GPU\u5b9e\u73b0330\u4e07Token\u4e0a\u4e0b\u6587\u63a8\u7406"}]},{"@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\/36299","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=36299"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36299\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36299"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36299"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36299"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}