{"id":748539,"date":"2026-03-18T11:13:07","date_gmt":"2026-03-18T03:13:07","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=748539"},"modified":"2026-03-18T11:13:07","modified_gmt":"2026-03-18T03:13:07","slug":"%e6%9d%a8%e6%a4%8d%e9%ba%9f%e5%b8%a6%e9%98%9f%e5%8f%91%e8%ae%ba%e6%96%87%ef%bc%8ckimi%e4%b8%8b%e4%b8%80%e4%bb%a3%e6%a8%a1%e5%9e%8b%e6%9e%b6%e6%9e%84%e6%9b%9d%e5%85%89%ef%bc%8c%e6%8e%a8%e7%90%86","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=748539","title":{"rendered":"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86"},"content":{"rendered":"<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/BAEF6BCAD3B8863679B22CFBA55C98D847C82C84_size12_w640_h272.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p><strong>\u667a\u4e1c\u897f<\/strong><\/p>\n<p><strong>\u4f5c\u8005\uff5c\u9648\u9a8f\u8fbe<\/strong><\/p>\n<p><strong>\u7f16\u8f91\uff5c\u4e91\u9e4f<\/strong><\/p>\n<p><strong>\u201c\u4e00\u5207\u90fd\u9700\u8981\u88ab\u91cd\u65b0\u601d\u8003\uff0c\u6df1\u5ea6\u5b66\u4e602.0\u65f6\u4ee3\u5373\u5c06\u6765\u4e34\u3002\u201d<\/strong>\u5728\u8bfb\u5b8c\u4e0b\u65b9\u8fd9\u7bc7\u6765\u81ea\u6708\u4e4b\u6697\u9762\u7684\u6700\u65b0\u8bba\u6587\u540e\uff0c\u524d<a href=\"https:\/\/www.alphaseek.cn\/?tag=openai\" title=\"View all posts in OpenAI\" target=\"_blank\" class=\"tag_link\">OpenAI<\/a>\u5927\u725b\u3001\u201c\u63a8\u7406\u6a21\u578b\u4e4b\u7236\u201dJerry Tworek\u53d1\u51fa\u611f\u53f9\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/6B0444D22DDF02869C3C2B54AAA18D257C9AD0F0_size56_w640_h659.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p>\u667a\u4e1c\u897f3\u670816\u65e5\u62a5\u9053\uff0c\u4eca\u5929\uff0c\u6708\u4e4b\u6697\u9762\u53d1\u5e03\u8bba\u6587\uff0c\u63d0\u524d\u9884\u89c8\u4e86\u4e0b\u4e00\u4ee3\u6a21\u578b\u7684\u5173\u952e\u6a21\u5757\u2014\u2014<strong>\u6ce8\u610f\u529b\u6b8b\u5dee\uff08Attention Residuals\uff0c\u7b80\u79f0AttnRes\uff09<\/strong>\u3002\u8bba\u6587\u7684\u6838\u5fc3\u4eae\u70b9\u5728\u4e8e\u5bf9\u5927\u6a21\u578b\u4e2d\u6700\u57fa\u7840\u3001\u4f46\u957f\u671f\u88ab\u5ffd\u89c6\u7684\u7ed3\u6784\u2014\u2014<strong>\u6b8b\u5dee\u8fde\u63a5\uff08Residual Connection\uff09<\/strong>\u8fdb\u884c\u4e86\u91cd\u65b0\u8bbe\u8ba1\u3002<\/p>\n<p>\u5728\u4f20\u7edfTransformer\u67b6\u6784\u4e2d\uff0c\u6bcf\u4e00\u5c42\u90fd\u4f1a\u5c06\u524d\u4e00\u5c42\u7684\u8f93\u51fa\u4e0e\u5f53\u524d\u5c42\u7684\u8ba1\u7b97\u7ed3\u679c\u76f4\u63a5\u76f8\u52a0\uff0c\u8fd9\u79cd\u201c\u7b49\u6743\u7d2f\u52a0\u201d\u7684\u65b9\u5f0f\u867d\u7136\u7a33\u5b9a\u4e86\u6df1\u5ea6\u7f51\u7edc\u8bad\u7ec3\uff0c\u4f46\u968f\u7740\u6a21\u578b\u5c42\u6570\u4e0d\u65ad\u589e\u52a0\uff0c\u6240\u6709\u5386\u53f2\u5c42\u7684\u8868\u793a\u88ab\u7b80\u5355\u53e0\u52a0\uff0c<strong>\u5bb9\u6613\u5bfc\u81f4\u4fe1\u606f\u88ab\u9010\u6e10\u7a00\u91ca\uff0c\u6df1\u5c42\u7f51\u7edc\u7684\u8868\u8fbe\u80fd\u529b\u53d7\u5230\u9650\u5236\u3002<\/strong><\/p>\n<p><strong>AttnRes\u7528\u6ce8\u610f\u529b\u673a\u5236\u66ff\u4ee3\u56fa\u5b9a\u7684\u6b8b\u5dee\u76f8\u52a0\u65b9\u5f0f<\/strong>\uff0c\u8ba9\u6bcf\u4e00\u5c42\u80fd\u591f\u4ece\u6240\u6709\u5386\u53f2\u5c42\u8868\u793a\u4e2d\u52a8\u6001\u9009\u62e9\u6700\u6709\u7528\u7684\u4fe1\u606f\u3002\u6bcf\u4e00\u5c42\u76f8\u5f53\u4e8e\u4e00\u4e2a\u67e5\u8be2\uff08Query\uff09\uff0c\u5728\u4e4b\u524d\u5c42\u7684\u8868\u793a\u4e2d\u8fdb\u884c\u68c0\u7d22\u5e76\u5206\u914d\u6743\u91cd\uff0c\u4f7f\u6a21\u578b\u80fd\u66f4\u7075\u6d3b\u5730\u5229\u7528\u4e0d\u540c\u5c42\u7ea7\u7684\u7279\u5f81\u3002<\/p>\n<p>\u5728\u5b9e\u73b0\u5c42\u9762\uff0c\u8bba\u6587\u8fd8\u63d0\u51fa\u4e86<strong>\u5757\u7ea7\u6ce8\u610f\u529b\u6b8b\u5dee\uff08Block AttnRes\uff09<\/strong>\u6765\u89e3\u51b3\u5927\u6a21\u578b\u8bad\u7ec3\u65f6\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u5f00\u9500\u95ee\u9898\u3002<\/p>\n<p>\u5b9e\u9a8c\u4e2d\uff0c\u6708\u4e4b\u6697\u9762\u5c06AttnRes\u96c6\u6210\u5230Kimi Linear\u67b6\u6784\uff08\u603b\u53c2\u6570\u91cf48B\/\u6fc0\u6d3b\u53c2\u6570\u91cf3B\uff09\u4e2d\uff0c\u57281.4T token\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u3002AttnRes\u7f13\u89e3\u4e86\u9884\u5f52\u4e00\u5316\u7684\u7a00\u91ca\u95ee\u9898\uff0c\u4ea7\u751f\u66f4\u5747\u5300\u7684\u8de8\u6df1\u5ea6\u8f93\u51fa\u5e45\u5ea6\u548c\u68af\u5ea6\u5206\u5e03\uff0c\u4ece\u800c\u7f13\u89e3\u6df1\u5c42\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u4e0d\u7a33\u5b9a\u95ee\u9898\u3002<\/p>\n<p>\u5728\u76f8\u540c\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u5f15\u5165AttnRes\u7684\u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5747\u53d6\u5f97\u7a33\u5b9a\u63d0\u5347\uff0c<strong>\u5c24\u5176\u662f\u5728\u6570\u5b66\u63a8\u7406\u3001\u79d1\u5b66\u95ee\u7b54\u4ee5\u53ca\u4ee3\u7801\u751f\u6210\u7b49\u9700\u8981\u590d\u6742\u63a8\u7406\u80fd\u529b\u7684\u4efb\u52a1\u4e0a\u3002<\/strong><\/p>\n<p>\u6b64\u5916\uff0cBlock AttnRes\u7684\u9a8c\u8bc1\u635f\u5931\u4e3a1.692\uff0c\u800c\u57fa\u7ebf\u4e3a1.714\uff0c<strong>\u76f8\u5f53\u4e8e\u7ea61.25\u500d\u7684\u8ba1\u7b97\u6548\u7387\u4f18\u52bf\u3002<\/strong><\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/49C9E3C0011AF5794625D1003A4834E225F4CFFF_size20_w500_h442.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u5728\u5927\u6a21\u578b\u5708\u5f15\u53d1\u5e7f\u6cdb\u8ba8\u8bba\uff0c\u6708\u4e4b\u6697\u9762\u5b98\u65b9\u5728X\u5e73\u53f0\u53d1\u5e03\u7684\u5e16\u5b50\u9605\u8bfb\u91cf\u5df2\u903c\u8fd160\u4e07\u6b21\u3002\u9664\u4e86Jerry Tworek\u4e4b\u5916\uff0cOpenAI\u8054\u5408\u521b\u59cb\u4ebaAndrej Karpathy\u4e5f\u8bc4\u8bba\u9053\uff1a\u201c<strong>\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff08\u5e38\u7528\u7684\u4f18\u5316\u7b97\u6cd5\u4e4b\u4e00\uff09\u4e5f\u662f\u4e00\u79cdResNet\uff0c\u6b8b\u5dee\u6d41\u5c31\u662f\u6743\u91cd\u2026\u2026\u770b\u6765\u6211\u4eec\u8fd8\u6ca1\u628a\u2018Attention is All You Need\u2019 \u8fd9\u53e5\u8bdd\u5b8c\u5168\u6309\u5b57\u9762\u610f\u601d\u7406\u89e3\u900f\u3002<\/strong>\u201d<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/CDA166737EB85A0398B3ACF754782A2A1640E5C6_size32_w500_h407.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p>\u6708\u4e4b\u6697\u9762\u7684\u6570\u5341\u540d\u7814\u7a76\u5458\u53c2\u4e0e\u4e86\u8fd9\u9879\u7814\u7a76\uff0c\u5bf9\u9879\u76ee\u8d21\u732e\u6700\u5927\u7684\u4e09\u4f4d\u7814\u7a76\u5458\u5305\u62ecGuangyu Chen\u3001Yu Zhang\uff08Kimi-Linear\u4e00\u4f5c\uff09\u3001Jialin Su\u3002\u6708\u4e4b\u6697\u9762\u7684\u4e09\u4f4d\u8054\u5408\u521b\u59cb\u4eba\u6768\u690d\u9e9f\u3001\u5434\u80b2\u6615\u3001\u5468\u6615\u5b87\u5e26\u9886\u56e2\u961f\u5b8c\u6210\u4e86\u8fd9\u9879\u7814\u7a76\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/F8EF09F095EF5BE35B1B213C16DE932F5EC19DB1_size24_w640_h371.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p><strong>\u4e00\u3001\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\uff0c\u9762\u4e34\u201c\u7b49\u6743\u7d2f\u52a0\u201d\u4e4b\u56f0<\/strong><\/p>\n<p>\u7531\u77e5\u540d\u534e\u4ebaAI\u5b66\u8005\u4f55\u607a\u660e\u7b49\u4eba\u63d0\u51fa\u7684\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\uff0c\u6700\u65e9\u5728ResNet\u4e2d\u7cfb\u7edf\u5316\u63d0\u51fa\uff0c\u5e76\u8fc5\u901f\u6210\u4e3a\u73b0\u4ee3\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u7684\u6838\u5fc3\u7ed3\u6784\u4e4b\u4e00\u3002<\/p>\n<p>\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0c\u8fd9\u4e00\u8bbe\u8ba1\u4e0d\u4ec5\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u53d6\u5f97\u5de8\u5927\u6210\u529f\uff0c\u4e5f\u9010\u6e10\u6210\u4e3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u548c\u5404\u7c7bTransformer\u7cfb\u7edf\u4e2d\u7684\u57fa\u7840\u6784\u5efa\u6a21\u5757\u3002<\/p>\n<p>\u5176\u6838\u5fc3\u601d\u60f3\u975e\u5e38\u76f4\u89c2\uff1a<strong>\u8ba9\u7f51\u7edc\u5b66\u4e60\u201c\u6b8b\u5dee\u201d\uff08\u5373\u8f93\u5165\u4e0e\u8f93\u51fa\u4e4b\u95f4\u7684\u5dee\u503c\uff09\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u62df\u5408\u5b8c\u6574\u6620\u5c04\u3002<\/strong>\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\u4e2d\uff0c\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\u7b49\u4e8e\u6240\u6709\u4e4b\u524d\u5c42\u8f93\u51fa\u7684\u7d2f\u79ef\u548c\u3002\u6b8b\u5dee\u8fde\u63a5\u5b9e\u9645\u4e0a\u662f\u4e00\u79cd\u8de8\u6df1\u5ea6\u7684\u4fe1\u606f\u805a\u5408\u673a\u5236\uff0c<strong>\u6240\u6709\u5c42\u7684\u8d21\u732e\u6743\u91cd\u90fd\u662f1\uff0c\u4e5f\u5c31\u662f\u5747\u5300\u52a0\u6743\u3002<\/strong><\/p>\n<p>\u5728\u73b0\u4ee3\u5927\u6a21\u578b\u4e2d\uff0c\u6b8b\u5dee\u7ed3\u6784\u901a\u5e38\u8fd8\u4f1a\u4e0e<strong>PreNorm\uff08\u9884\u5f52\u4e00\u5316\uff09<\/strong>\u7ed3\u5408\u4f7f\u7528\u3002\u5177\u4f53\u505a\u6cd5\u662f\u5728\u6bcf\u4e2a\u5b50\u5c42\u8ba1\u7b97\u4e4b\u524d\u5148\u8fdb\u884c\u5c42\u5f52\u4e00\u5316\uff0c\u7136\u540e\u518d\u8fdb\u5165\u6ce8\u610f\u529b\u6216\u524d\u9988\u7f51\u7edc\u6a21\u5757\uff0c\u6700\u540e\u518d\u901a\u8fc7\u6b8b\u5dee\u8fde\u63a5\u5c06\u8f93\u51fa\u52a0\u56de\u539f\u59cb\u8868\u793a\u3002<\/p>\n<p>\u8fd9\u79cd\u7ed3\u6784\u5728Transformer\u6a21\u578b\u4e2d\u88ab\u5e7f\u6cdb\u91c7\u7528\uff0c\u56e0\u4e3a\u5b83\u80fd\u591f\u8fdb\u4e00\u6b65\u6539\u5584\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u4f7f\u68af\u5ea6\u5728\u6df1\u5c42\u7f51\u7edc\u4e2d\u66f4\u5bb9\u6613\u4f20\u64ad\uff0c\u5e76\u907f\u514d\u8bad\u7ec3\u540e\u671f\u51fa\u73b0\u6570\u503c\u4e0d\u7a33\u5b9a\u7684\u95ee\u9898\u3002<\/p>\n<p>\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\u4e4b\u6240\u4ee5\u80fd\u83b7\u5f97\u5e7f\u6cdb\u91c7\u7528\uff0c\u5728\u4e8e\u5176\u89e3\u51b3\u4e86\u6df1\u5c42\u7f51\u7edc\u8bad\u7ec3\u56f0\u96be\u7684\u95ee\u9898\uff0c\u5141\u8bb8\u68af\u5ea6\u7ed5\u8fc7\u590d\u6742\u53d8\u6362\uff0c\u7f13\u89e3\u4e86\u68af\u5ea6\u6d88\u5931\u7684\u95ee\u9898\u3002\u540c\u65f6\uff0c\u6b8b\u5dee\u4f7f\u6bcf\u5c42\u5bf9\u8868\u793a\u7684\u4fee\u6539\u662f\u6e10\u8fdb\u7684\uff0c\u4e0d\u4f1a\u7834\u574f\u5df2\u6709\u8868\u793a\u3002\u6a21\u578b\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3001\u66f4\u5bb9\u6613\u4f18\u5316\u3002<\/p>\n<p>\u7136\u800c\uff0c\u8fd9\u4e00\u65b9\u5f0f\u4e5f\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u6027\u3002\u6807\u51c6\u6b8b\u5dee\u8fde\u63a5\u4e2d\u6240\u6709\u5c42\u7684\u6743\u91cd\u4e00\u81f4\uff0c\u4f46\u73b0\u5b9e\u60c5\u51b5\u662f\u4e0d\u540c\u5c42\u4fe1\u606f\u7684\u91cd\u8981\u6027\u4e0d\u540c\u3002<strong>\u7406\u60f3\u60c5\u51b5\u4e0b\u5e94\u8be5\u91c7\u7528\u53ef\u5b66\u4e60\u7684\u6743\u91cd\u3002<\/strong><\/p>\n<p>\u6b64\u5916\uff0c\u5b66\u4e60\u65f6\uff0c\u6bcf\u5c42\u7f51\u7edc\u53ea\u80fd\u770b\u5230\u538b\u7f29\u540e\u7684\u72b6\u6001\uff0c\u5bfc\u81f4\u8bb8\u591a\u65e9\u671f\u4fe1\u606f\u88ab\u6df7\u5408\u538b\u7f29\uff0c\u65e0\u6cd5\u5355\u72ec\u8bbf\u95ee\u67d0\u4e00\u5c42\u8f93\u51fa\uff0c\u5bfc\u81f4\u4fe1\u606f\u4e0d\u53ef\u9006\u7684\u4e22\u5931\u3002<strong>\u540c\u65f6\uff0c\u540e\u671f\u5c42\u5728\u6a21\u578b\u7684\u8d21\u732e\u4e2d\u53d8\u5927\uff0c\u65e9\u671f\u5c42\u88ab\u9010\u6e10\u6df9\u6ca1\u3002<\/strong><\/p>\n<p>\u6a21\u578b\u526a\u679d\uff08prune\uff09\u7b49\u65b9\u5f0f\u5bf9\u6a21\u578b\u6027\u80fd\u5f71\u54cd\u4e0d\u5927\uff0c\u8fd9\u4e00\u73b0\u8c61\u4e5f\u4ece\u4fa7\u9762\u8bc1\u660e\uff0c<strong>\u5f53\u524d\u5e7f\u6cdb\u91c7\u7528\u7684\u6b8b\u5dee\u7ed3\u6784\uff0c\u4f7f\u5f88\u591a\u5c42\u8d21\u732e\u5f88\u5c0f\u3002<\/strong><\/p>\n<p>\u6b63\u662f\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\uff0c\u6708\u4e4b\u6697\u9762\u7684\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u601d\u8def\uff1a\u7528\u4e60\u5f97\u7684softmax\u6ce8\u610f\u529b\u673a\u5236\u6765\u66ff\u4ee3\u56fa\u5b9a\u7684\u6b8b\u5dee\u7d2f\u79ef\uff0c\u4e5f\u5c31\u662fAttnRes\u3002<\/p>\n<p><strong>\u4e8c\u3001\u7528\u6ce8\u610f\u529b\u66ff\u4ee3\u56fa\u5b9a\u6743\u91cd\uff0c\u8ba9\u6bcf\u4e00\u5c42\u5b66\u4f1a\u201c\u68c0\u7d22\u201d\u5386\u53f2<\/strong><\/p>\n<p>\u5728AttnRes\u4e2d\uff0c\u6bcf\u4e00\u5c42\u5728\u8ba1\u7b97\u65f6\u4e0d\u518d\u4ec5\u4ec5\u4f9d\u8d56\u4e0a\u4e00\u5c42\u7684\u8f93\u51fa\uff0c\u800c\u662f\u53ef\u4ee5\u5bf9\u4e4b\u524d\u591a\u4e2a\u5c42\u7684\u8868\u793a\u8fdb\u884c\u4e00\u6b21\u6ce8\u610f\u529b\u68c0\u7d22\u3002\u6a21\u578b\u901a\u8fc7\u5b66\u4e60\u5f97\u5230\u7684\u4e00\u7ec4\u6743\u91cd\uff0c\u5bf9\u4e0d\u540c\u5c42\u7684\u8f93\u51fa\u8fdb\u884c\u52a0\u6743\u7ec4\u5408\uff0c\u4ece\u800c\u751f\u6210\u5f53\u524d\u5c42\u7684\u8f93\u5165\u8868\u793a\u3002<\/p>\n<p>\u7531\u4e8e\u8fd9\u4e9b\u6743\u91cd\u901a\u8fc7softmax\u5f52\u4e00\u5316\uff0c<strong>\u5b83\u4eec\u80fd\u591f\u53cd\u6620\u4e0d\u540c\u5c42\u5728\u5f53\u524d\u4e0a\u4e0b\u6587\u4e2d\u7684\u76f8\u5bf9\u91cd\u8981\u6027\uff0c\u4f7f\u4fe1\u606f\u805a\u5408\u8fc7\u7a0b\u66f4\u52a0\u7075\u6d3b\u548c\u81ea\u9002\u5e94\u3002<\/strong><\/p>\n<p>\u8fd9\u79cd\u673a\u5236\u5e26\u6765\u7684\u4e00\u4e2a\u76f4\u63a5\u597d\u5904\u662f\uff0c\u6a21\u578b\u80fd\u591f\u66f4\u6709\u6548\u5730\u5229\u7528\u65e9\u671f\u5c42\u7684\u4fe1\u606f\u3002\u7531\u4e8e\u6bcf\u4e00\u5c42\u90fd\u53ef\u4ee5\u76f4\u63a5\u8bbf\u95ee\u5386\u53f2\u5c42\u7684\u8868\u793a\uff0c\u65e9\u671f\u4fe1\u606f\u4e0d\u518d\u88ab\u5b8c\u5168\u538b\u7f29\u5230\u5355\u4e00\u72b6\u6001\u4e2d\uff0c\u800c\u662f\u53ef\u4ee5\u5728\u9700\u8981\u65f6\u88ab\u91cd\u65b0\u8c03\u7528\u3002<\/p>\n<p>\u8fd9\u79cd\u201c\u8de8\u5c42\u68c0\u7d22\u201d\u7684\u80fd\u529b\uff0c\u4f7f\u5f97\u6df1\u5c42\u7f51\u7edc\u5728\u4fe1\u606f\u5229\u7528\u4e0a\u66f4\u52a0\u63a5\u8fd1\u4e8eTransformer\u5728\u65f6\u95f4\u7ef4\u5ea6\u4e0a\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2atoken\uff0c\u5b8c\u6574AttnRes\u9700\u8981O(L&sup2;d)\u7684\u7b97\u672f\u8fd0\u7b97\u548cO(Ld)\u7684\u5185\u5b58\u6765\u5b58\u50a8\u5c42\u8f93\u51fa\u3002<strong>\u7531\u4e8e\u6df1\u5ea6\u8fdc\u5c0f\u4e8e\u5e8f\u5217\u957f\u5ea6\uff0c\u7b97\u672f\u6210\u672c\u9002\u4e2d\u3002<\/strong><\/p>\n<p>\u5f53\u7136\uff0c\u5982\u679c\u8ba9\u6bcf\u4e00\u5c42\u90fd\u5bf9\u6240\u6709\u5386\u53f2\u5c42\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u4e5f\u5c31\u662f\u4f7f\u7528<strong>\u5168\u6ce8\u610f\u529b\u6b8b\u5dee<\/strong>\uff0c\u8ba1\u7b97\u6210\u672c\u4f1a\u8fc5\u901f\u589e\u52a0\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u56e2\u961f\u8fdb\u4e00\u6b65\u63d0\u51fa\u4e86<strong>Block AttnRes\uff08\u5757\u6ce8\u610f\u529b\u6b8b\u5dee\uff09<\/strong>\u7684\u7ed3\u6784\u8bbe\u8ba1\u3002<\/p>\n<p>\u8be5\u65b9\u6cd5\u5c06\u7f51\u7edc\u5c42\u5212\u5206\u4e3a\u82e5\u5e72\u5757\uff0c\u6bcf\u4e2a\u5757\u5185\u90e8\u4fdd\u6301\u5e38\u89c4\u8ba1\u7b97\u65b9\u5f0f\uff0c\u800c\u8de8\u5757\u4e4b\u95f4\u5219\u4f7f\u7528\u538b\u7f29\u540e\u7684\u8868\u793a\u8fdb\u884c\u6ce8\u610f\u529b\u4ea4\u4e92\u3002<\/p>\n<p>\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c<strong>\u6a21\u578b\u5728\u4fdd\u6301\u7075\u6d3b\u4fe1\u606f\u68c0\u7d22\u80fd\u529b\u7684\u540c\u65f6\uff0c\u4e5f\u80fd\u6709\u6548\u63a7\u5236\u8ba1\u7b97\u590d\u6742\u5ea6<\/strong>\uff0c\u4f7f\u8fd9\u4e00\u673a\u5236\u5728\u5927\u89c4\u6a21\u6a21\u578b\u4e2d\u4ecd\u7136\u5177\u6709\u53ef\u884c\u6027\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/E6C555E040A5513E110D4D406B1531E46DFB2B72_size42_w640_h370.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u6768\u690d\u9e9f\u5e26\u961f\u53d1\u8bba\u6587\uff0cKimi\u4e0b\u4e00\u4ee3\u6a21\u578b\u67b6\u6784\u66dd\u5149\uff0c\u63a8\u7406\u4e4b\u7236\u60ca\u547c\uff1a\u6df1\u5ea6\u5b66\u4e602.0\u8981\u6765\u4e86\"  \/><\/p>\n<p>\u4ece\u5de6\u5230\u53f3\uff1a\u4f20\u7edf\u6b8b\u5dee\u3001\u5168\u6ce8\u610f\u529b\u6b8b\u5dee\u3001\u5757\u6ce8\u610f\u529b\u6b8b\u5dee<\/p>\n<p><strong>\u4e09\u3001\u57fa\u7840\u8bbe\u65bd\u534f\u540c\u4f18\u5316\uff0c\u8ba1\u7b97\u6548\u7387\u8fbe1.25\u500d<\/strong><\/p>\n<p>\u4e0d\u8fc7\uff0cBlock AttnRes\u4e5f\u7ed9\u57fa\u7840\u8bbe\u65bd\u5e26\u6765\u65b0\u7684\u6311\u6218\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u9636\u6bb5\uff0cBlock AttnRes\u5e26\u6765\u66f4\u591a\u7684<strong>\u6d41\u6c34\u7ebf\u5e76\u884c\u4e2d\u901a\u4fe1\u5f00\u9500<\/strong>\u3002\u6807\u51c6\u6b8b\u5dee\u7ed3\u6784\u5728\u6d41\u6c34\u7ebf\u9636\u6bb5\u4e4b\u95f4\u53ea\u9700\u4f20\u8f93\u56fa\u5b9a\u5927\u5c0f\u7684\u9690\u85cf\u72b6\u6001\uff0c\u800cBlock AttnRes\u9700\u8981\u5728\u6bcf\u4e2a\u9636\u6bb5\u8bbf\u95ee\u6b64\u524d\u6240\u6709\u5757\u7684\u8868\u5f81\u3002<\/p>\n<p>\u5982\u679c\u91c7\u7528\u6734\u7d20\u5b9e\u73b0\uff0c\u5c31\u5fc5\u987b\u5728\u9636\u6bb5\u8f6c\u6362\u65f6\u91cd\u590d\u4f20\u8f93\u5b8c\u6574\u7684\u5386\u53f2\u5757\u8868\u793a\uff0c\u901a\u4fe1\u91cf\u4f1a\u968f\u7740\u5757\u6570\u91cf\u7d2f\u79ef\u5448\u5e73\u65b9\u589e\u957f\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6708\u4e4b\u6697\u9762\u5f15\u5165<strong>\u8de8\u9636\u6bb5\u7f13\u5b58\uff08cross-stage cache\uff09<\/strong>\u673a\u5236\uff0c\u5927\u5e45\u51cf\u5c11\u5197\u4f59\u901a\u4fe1\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u6bcf\u4e2a\u5757\u53ea\u9700\u5728\u865a\u62df\u9636\u6bb5\u4e2d\u5b58\u50a8\u4e00\u6b21\uff0c\u7ed3\u5408\u6fc0\u6d3b\u68c0\u67e5\u70b9\u673a\u5236\u540e\uff0c\u6bcf\u5c42\u6fc0\u6d3b\u5185\u5b58\u5360\u7528\u4e0e\u6807\u51c6Transformer\u57fa\u672c\u4e00\u81f4\uff0c\u56e0\u6b64\u6574\u4f53\u8bad\u7ec3\u7684\u989d\u5916\u5f00\u9500\u5f88\u5c0f\uff0c\u5728\u5b9e\u9645\u7cfb\u7edf\u4e2d\u7aef\u5230\u7aef\u8bad\u7ec3\u65f6\u95f4\u589e\u52a0\u4e0d\u52304%\u3002<\/p>\n<p>\u5728\u63a8\u7406\u9636\u6bb5\uff0cBlock AttnRes\u7684\u4e3b\u8981\u6311\u6218\u6765\u81ea<strong>\u8de8\u5c42\u5757\u8868\u5f81\u7684\u91cd\u590d\u8bbf\u95ee\u548c\u957f\u4e0a\u4e0b\u6587\u7f13\u5b58\u5e26\u6765\u7684\u5185\u5b58\u538b\u529b<\/strong>\u3002\u4e3a\u6b64\u4ed6\u4eec\u91c7\u7528\u4e86<strong>\u4e24\u9636\u6bb5\u8ba1\u7b97<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e7%ad%96%e7%95%a5\" title=\"View all posts in \u7b56\u7565\" target=\"_blank\" class=\"tag_link\">\u7b56\u7565<\/a>\uff0c<\/strong>\u9996\u5148\u5728\u9636\u6bb5\u4e00\u4e2d\u5bf9\u7f13\u5b58\u7684\u5757\u8868\u5f81\u6267\u884c\u4e00\u6b21\u6279\u91cf\u67e5\u8be2\uff0c\u4e3a\u540c\u4e00\u5757\u5185\u7684\u6240\u6709\u5c42\u540c\u65f6\u8ba1\u7b97\u5757\u95f4\u6ce8\u610f\u529b\uff0c\u4ece\u800c\u5c06\u539f\u672c\u6bcf\u5c42\u90fd\u9700\u8981\u8bbf\u95ee\u5386\u53f2\u5757\u7684\u64cd\u4f5c\u5408\u5e76\u4e3a\u4e00\u6b21\u77e9\u9635\u8ba1\u7b97\uff0c\u663e\u8457\u964d\u4f4e\u5185\u5b58\u8bfb\u53d6\u6b21\u6570\u3002<\/p>\n<p>\u968f\u540e\u5728\u9636\u6bb5\u4e8c\u4e2d\u6309\u987a\u5e8f\u8ba1\u7b97\u5757\u5185\u5404\u5c42\u7684\u6ce8\u610f\u529b\uff0c\u5e76\u901a\u8fc7\u5728\u7ebfsoftmax\u5c06\u7ed3\u679c\u4e0e\u9636\u6bb5\u4e00\u7684\u8f93\u51fa\u5408\u5e76\uff0c\u8fd9\u4e00\u9636\u6bb5\u53ef\u4ee5\u4e0e\u5468\u56f4\u7b97\u5b50\u8fdb\u884c\u5185\u6838\u878d\u5408\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u51cf\u5c11I\/O\u5f00\u9500\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4e3a\u7f13\u89e3\u957f\u4e0a\u4e0b\u6587\u9884\u586b\u5145\u65f6\u5757\u8868\u5f81\u7f13\u5b58\u8fc7\u5927\u7684\u95ee\u9898\uff0c\u7cfb\u7edf\u5c06\u8fd9\u4e9b\u8868\u5f81\u6cbf\u5e8f\u5217\u7ef4\u5ea6\u5728\u5f20\u91cf\u5e76\u884c\u8bbe\u5907\u4e4b\u95f4\u8fdb\u884c\u5206\u7247\uff0c\u4f7f\u6bcf\u4e2a\u8bbe\u5907\u53ea\u5b58\u50a8\u90e8\u5206\u5e8f\u5217\u3002\u901a\u8fc7\u4e24\u9636\u6bb5\u8ba1\u7b97\u548c\u5e8f\u5217\u5206\u7247\u7684\u7ed3\u5408\uff0c<strong>BlockAttnRes\u5728\u63a8\u7406\u65f6\u4ec5\u5e26\u6765\u4e0d\u52302%\u7684\u5ef6\u8fdf\u5f00\u9500\uff0c\u540c\u65f6\u663e\u8457\u964d\u4f4e\u4e86\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e0b\u7684\u663e\u5b58\u9700\u6c42\u3002<\/strong><\/p>\n<p>\u5728\u5b9e\u9a8c\u65b9\u9762\uff0c\u7814\u7a76\u56e2\u961f\u5c06\u8fd9\u4e00\u7ed3\u6784\u96c6\u6210\u5230Kimi Linear\u67b6\u6784\u4e2d\u8fdb\u884c\u9a8c\u8bc1\u3002<\/p>\n<p>\u5b9e\u9a8c\u8868\u660e\uff0c\u5728\u76f8\u540c\u53c2\u6570\u89c4\u6a21\u4e0b\uff0c\u5f15\u5165Block AttnRes\u540e\uff0c\u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u53d6\u5f97\u4e86\u6027\u80fd\u63d0\u5347\u3002\u5176\u4e2d\u63d0\u5347\u6700\u591a\u7684\u662f\u591a\u6b65\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5GPQA-Diamond\uff0c<strong>\u91c7\u7528AttnRes\u7684\u6a21\u578b\u5f97\u5206\u63d0\u5347\u5e45\u5ea6\u8d85\u8fc7\u4e8620%\u3002<\/strong><\/p>\n<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u8be5\u65b9\u6cd5\u8fd8\u8868\u73b0\u51fa\u826f\u597d\u7684\u8ba1\u7b97\u6548\u7387\uff0cBlock AttnRes\u5728\u635f\u5931\u51fd\u6570\u4e0a\u4e0e\u4f7f\u75281.25\u500d\u8ba1\u7b97\u91cf\u8bad\u7ec3\u7684\u57fa\u7ebf\u6a21\u578b\u76f8\u5f53\u3002<\/p>\n<p><strong>\u7ed3\u8bed\uff1a\u5c06\u6ce8\u610f\u529b\u5f15\u5165\u201c\u7f51\u7edc\u6df1\u5ea6\u7ef4\u5ea6\u201d\uff0c\u6709\u671b\u6210\u4e3a\u4e0b\u4e00\u4ee3\u6a21\u578b\u5173\u952e\u67b6\u6784<\/strong><\/p>\n<p>\u901a\u8fc7\u5c06\u6ce8\u610f\u529b\u673a\u5236\u5f15\u5165\u6df1\u5ea6\u7ef4\u5ea6\uff0cAttention Residuals\u4e3a\u6df1\u5c42\u7f51\u7edc\u63d0\u4f9b\u4e86\u4e00\u79cd\u66f4\u52a0\u7075\u6d3b\u7684\u4fe1\u606f\u805a\u5408\u65b9\u5f0f\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u52a8\u6001\u9009\u62e9\u5386\u53f2\u8868\u793a\uff0c\u800c\u4e0d\u662f\u88ab\u52a8\u63a5\u53d7\u5747\u5300\u7d2f\u79ef\u7684\u7ed3\u679c\u3002\u8fd9\u4e00\u601d\u8def\u6709\u671b\u8fdb\u4e00\u6b65\u63d0\u5347\u6df1\u5c42\u6a21\u578b\u7684\u4fe1\u606f\u5229\u7528\u6548\u7387\u3002<\/p>\n<p>\u4ece\u67b6\u6784\u6f14\u8fdb\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u8fd9\u9879\u5de5\u4f5c\u4e5f\u63d0\u51fa\u4e86\u4e00\u4e2a\u503c\u5f97\u5173\u6ce8\u7684\u601d\u8def\uff1a\u5728\u8fc7\u53bb\uff0c\u6ce8\u610f\u529b\u673a\u5236\u4e3b\u8981\u7528\u4e8e\u89e3\u51b3\u5e8f\u5217\u7ef4\u5ea6\u4e0a\u7684\u4fe1\u606f\u4ea4\u4e92\u95ee\u9898\uff0c\u800cAttnRes\u5219\u5c06\u6ce8\u610f\u529b\u5f15\u5165\u5230\u4e86\u201c\u7f51\u7edc\u6df1\u5ea6\u7ef4\u5ea6\u201d\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u5728\u4e0d\u540c\u5c42\u4e4b\u95f4\u5efa\u7acb\u66f4\u52a0\u7075\u6d3b\u7684\u4fe1\u606f\u901a\u8def\uff0c\u53ef\u80fd\u4e3a\u4e0b\u4e00\u4ee3\u5927\u6a21\u578b\u67b6\u6784\u6f14\u8fdb\u63d0\u4f9b\u65b0\u7684\u65b9\u5411\u3002<\/p>\n<p>(\u6765\u6e90\uff1a\u65b0\u6d6a\u79d1\u6280)<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u667a\u4e1c\u897f \u4f5c\u8005\uff5c\u9648\u9a8f\u8fbe \u7f16\u8f91\uff5c\u4e91\u9e4f \u201c\u4e00\u5207\u90fd\u9700\u8981\u88ab\u91cd\u65b0\u601d\u8003\uff0c\u6df1\u5ea6\u5b66\u4e602.0\u65f6\u4ee3\u5373\u5c06\u6765\u4e34\u3002\u201d\u5728\u8bfb\u5b8c\u4e0b\u65b9\u8fd9\u7bc7\u6765\u81ea\u6708\u4e4b [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[378,88],"views":31,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/748539"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=748539"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/748539\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=748539"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=748539"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=748539"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}