{"id":17231,"date":"2023-12-29T12:17:31","date_gmt":"2023-12-29T04:17:31","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=17231"},"modified":"2023-12-29T12:17:31","modified_gmt":"2023-12-29T04:17:31","slug":"%e8%b0%81%e8%83%bd%e6%92%bc%e5%8a%a8transformer%e7%bb%9f%e6%b2%bb%e5%9c%b0%e4%bd%8d%ef%bc%9fmamba%e4%bd%9c%e8%80%85%e8%b0%88llm%e6%9c%aa%e6%9d%a5%e6%9e%b6%e6%9e%84","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2023\/12\/29\/17231\/","title":{"rendered":"\u8c01\u80fd\u64bc\u52a8Transformer\u7edf\u6cbb\u5730\u4f4d\uff1fMamba\u4f5c\u8005\u8c08LLM\u672a\u6765\u67b6\u6784"},"content":{"rendered":"<div class=\"article-desc\">\u5728\u6700\u8fd1\u7684\u4e00\u6b21\u91c7\u8bbf\u4e2d\uff0c\u6765\u81ea\u827e\u4f26\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u6240\uff08 AI2\uff09\u7684\u673a\u5668\u5b66\u4e60\u7814\u7a76\u5458 Nathan Lambert \u4e0e Mamba \u8bba\u6587\u4f5c\u8005\u4e4b\u4e00 Tri Dao\uff0c\u4ee5\u53ca\u540c\u6837\u6765\u81ea Together.AI \u7684\u79d1\u5b66\u5bb6 Michael Poli \u8fdb\u884c\u4e86\u6df1\u5165\u4ea4\u6d41\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<p>\u5728\u5927\u6a21\u578b\u9886\u57df\uff0c\u4e00\u76f4\u7a33\u7ad9 C \u4f4d\u7684 Transformer \u6700\u8fd1\u4f3c\u4e4e\u6709\u88ab\u8d85\u8d8a\u7684\u8d8b\u52bf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u4e2a\u6311\u6218\u8005\u5c31\u662f\u4e00\u9879\u540d\u4e3a\u300cMamba\u300d\u7684\u7814\u7a76\uff0c\u5176\u5728\u8bed\u8a00\u3001\u97f3\u9891\u548c\u57fa\u56e0\u7ec4\u5b66\u7b49\u591a\u79cd\u6a21\u6001\u4e2d\u90fd\u8fbe\u5230\u4e86 SOTA \u6027\u80fd\u3002\u5728\u8bed\u8a00\u5efa\u6a21\u65b9\u9762\uff0c\u65e0\u8bba\u662f\u9884\u8bad\u7ec3\u8fd8\u662f\u4e0b\u6e38\u8bc4\u4f30\uff0cMamba-3B \u6a21\u578b\u90fd\u4f18\u4e8e\u540c\u7b49\u89c4\u6a21\u7684 Transformer \u6a21\u578b\uff0c\u5e76\u80fd\u4e0e\u4e24\u500d\u4e8e\u5176\u89c4\u6a21\u7684 Transformer \u6a21\u578b\u76f8\u5ab2\u7f8e\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img data-dominant-color=\"66514b\" data-has-transparency=\"false\" style=\"--dominant-color: #66514b;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-17232\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6.png\" width=\"1080\" height=\"617\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6.png 1080w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6-300x171.png 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6-1024x585.png 1024w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6-150x86.png 150w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2023\/12\/5589eea7300e580119d211ec738ca8ff20d8a6-768x439.png 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u8bba\u6587\u4e00\u7ecf\u53d1\u8868\uff0c\u5f15\u8d77\u4e86\u4e0d\u5c0f\u7684\u8f70\u52a8\u3002\u60ca\u53f9\u4e4b\u4f59\uff0c\u5927\u5bb6\u53d1\u73b0\u8bba\u6587\u4f5c\u8005\u53ea\u6709\u4e24\u4f4d\uff0c\u4e00\u4f4d\u662f\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u673a\u5668\u5b66\u4e60\u7cfb\u52a9\u7406\u6559\u6388 Albert Gu\uff0c\u53e6\u4e00\u4f4d\u662f Together.AI \u9996\u5e2d\u79d1\u5b66\u5bb6\u3001\u666e\u6797\u65af\u987f\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u52a9\u7406\u6559\u6388\uff08\u5373\u5c06\u4e0a\u4efb\uff09Tri Dao\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u9879\u7814\u7a76\u7684\u4e00\u4e2a\u91cd\u8981\u521b\u65b0\u662f\u5f15\u5165\u4e86\u4e00\u4e2a\u540d\u4e3a\u300c\u9009\u62e9\u6027 SSM\uff08 selective state space model\uff09\u300d\u7684\u67b6\u6784\uff0c\u76f8\u6bd4\u4e8e Transformer \u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u91cf\u4f1a\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u589e\u52a0\u5448\u5e73\u65b9\u7ea7\u589e\u957f\uff0c\u6bd4\u5982\u4e0a\u4e0b\u6587\u589e\u52a0 32 \u500d\u65f6\uff0c\u8ba1\u7b97\u91cf\u53ef\u80fd\u4f1a\u589e\u957f 1000 \u500d\uff0cMamba \u53ef\u4ee5\u968f\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u589e\u52a0\u5b9e\u73b0\u7ebf\u6027\u6269\u5c55\uff0c\u5176\u6027\u80fd\u5728\u5b9e\u9645\u6570\u636e\u4e2d\u53ef\u63d0\u9ad8\u5230\u767e\u4e07 token \u957f\u5ea6\u5e8f\u5217\uff0c\u5e76\u5b9e\u73b0 5 \u500d\u7684\u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347\u3002\u800c\u8fd9\u4e9b\uff0c\u90fd\u79bb\u4e0d\u5f00\u9009\u62e9\u6027 SSM\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u770b\u5230 Mamba \u5982\u6b64\u4f18\u5f02\u7684\u8868\u73b0\u540e\uff0c\u5f88\u591a\u7814\u7a76\u8005\u5bf9 SSM\uff08state space model\uff09 \u76f8\u5173\u7814\u7a76\u4ea7\u751f\u4e86\u597d\u5947\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5728\u6700\u8fd1\u7684\u4e00\u6b21\u91c7\u8bbf\u4e2d\uff0c\u6765\u81ea\u827e\u4f26\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u6240\uff08 AI2\uff09\u7684\u673a\u5668\u5b66\u4e60\u7814\u7a76\u5458\u00a0Nathan Lambert\u00a0\u4e0e Mamba \u8bba\u6587\u4f5c\u8005\u4e4b\u4e00\u00a0Tri Dao\uff0c\u4ee5\u53ca\u540c\u6837\u6765\u81ea Together.AI \u7684\u79d1\u5b66\u5bb6\u00a0Michael Poli\u00a0\u8fdb\u884c\u4e86\u6df1\u5165\u4ea4\u6d41\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4ed6\u4eec\u4e3b\u8981\u8ba8\u8bba\u4e86 LLM \u67b6\u6784\u7684\u672a\u6765\uff0c\u6b64\u5916\uff0c\u8fd9\u4e09\u4f4d\u7814\u7a76\u8005\u8fd8\u63a2\u8ba8\u4e86\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\uff08SSM\uff09\u5728\u65b0\u5174\u7684 LLM \u5e02\u573a\u4e2d\u7684\u5e94\u7528\u524d\u666f\u3002\u5bf9\u8bdd\u4e2d\u6d89\u53ca\u7684\u77e5\u8bc6\u70b9\u4e5f\u6bd4\u8f83\u5bc6\u96c6\uff0c\u6bd4\u5982\u4e3a\u4ec0\u4e48 Transformer \u4e2d\u7684\u6ce8\u610f\u529b\u673a\u5236\u6709\u6548\u3001\u5176\u6269\u5c55\u9650\u5236\u662f\u4ec0\u4e48\u3001Mamba \u4ecb\u7ecd\u53ca\u5176\u786c\u4ef6\u4f18\u5316\u4ee5\u53ca\u5bf9\u672a\u6765\u67b6\u6784\u9884\u6d4b\u7684\u8ba8\u8bba\u7b49\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4ee5\u4e0b\u662f\u5bf9\u8bdd\u5185\u5bb9\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>\u4e3a\u4ec0\u4e48\u6ce8\u610f\u529b\u673a\u5236\u6709\u6548<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u6211\u4eec\u9996\u5148\u6765\u8ba8\u8bba\u4e00\u4e0b\uff0c\u4e3a\u4ec0\u4e48\u6ce8\u610f\u529b\u673a\u5236\u6709\u6548\uff0c\u4ee5\u53ca\u6ce8\u610f\u529b\u673a\u5236\u7684\u5c40\u9650\u6027\u5728\u54ea\u91cc\u3002Transformer \u6709\u591a\u5c11\u90e8\u5206\u662f\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u6784\u5efa\u7684\uff0c\u8fd8\u6709\u6ca1\u6709\u5176\u4ed6\u7684\u673a\u5236\u5728\u8d77\u4f5c\u7528\uff0c\u5728\u8fd9\u65b9\u9762\u53ef\u80fd\u4f1a\u9047\u5230\u54ea\u4e9b\u6311\u6218\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>: \u6ca1\u9519\uff0c\u6240\u8c13 Transformer\uff0c\u5c31\u662f\u76ee\u524d\u63a8\u52a8\u6211\u4eec\u770b\u5230\u7684\u5927\u591a\u6570\u6fc0\u52a8\u4eba\u5fc3\u7684\u5e94\u7528\u7684\u67b6\u6784\u3002\u5982\u4f60\u6240\u8bf4\uff0c\u6ce8\u610f\u529b\u673a\u5236\u662f\u5176\u4e2d\u7684\u6838\u5fc3\u5c42\u3002\u5176\u5b9e\uff0c\u6ce8\u610f\u529b\u673a\u5236\u65e9\u5728 2014 \u5e74\u5230 2015 \u5e74\u5c31\u53d7\u5230\u5173\u6ce8\u4e86\uff0c\u968f\u540e Transformer \u7684\u6982\u5ff5\u51fa\u73b0\uff0c\u878d\u5408\u4e86\u6ce8\u610f\u529b\u673a\u5236\uff0c\u6ce8\u91cd\u591a\u5c42\u611f\u77e5\u673a\uff08MLP\uff09\u548c\u6ce8\u610f\u529b\u673a\u5236\u7684\u4ea4\u7ec7\u4f7f\u7528\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u8ba4\u4e3a\u5b83\u7684\u6210\u529f\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5728\u4e8e\uff0c\u8fd9\u4e9b\u6a21\u578b\u4f3c\u4e4e\u80fd\u591f\u5f88\u597d\u5730\u8fdb\u884c\u6269\u5c55\uff0c\u4f60\u53ef\u4ee5\u901a\u8fc7\u589e\u52a0\u66f4\u591a\u7684\u53c2\u6570\u548c\u6570\u636e\u6765\u6269\u5927\u6a21\u578b\u89c4\u6a21\u3002\u8fd9\u5c31\u662f\u6210\u529f\u7684\u79d8\u8bc0\u3002\u867d\u7136\u73b0\u5728\u770b\u8d77\u6765\u663e\u800c\u6613\u89c1\uff0c\u4f46\u6211\u8ba4\u4e3a\u4e94\u5e74\u524d\u8fd9\u8fd8\u4e0d\u662f\u4e00\u4e2a\u660e\u786e\u7684\u6982\u5ff5\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>Transformer \u6210\u529f\u7684\u51e0\u4e2a\u539f\u56e0\u5305\u62ec\uff1a\u9996\u5148\uff0c\u5b83\u8db3\u591f\u901a\u7528\uff0c\u80fd\u591f\u4ece\u5927\u91cf\u6570\u636e\u4e2d\u5b66\u5230\u5f88\u591a\u4e1c\u897f\u3002\u5176\u6b21\uff0c\u5b83\u5bf9\u786c\u4ef6\u975e\u5e38\u53cb\u597d\u3002\u4e0e\u4e4b\u524d\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u4e0d\u540c\uff0c\u5b83\u6ca1\u6709\u987a\u5e8f\u4f9d\u8d56\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u56e0\u6b64\uff0c\u5b83\u53ef\u4ee5\u5728 GPU\u3001TPU \u4e0a\u8fd0\u884c\u5f97\u5f88\u597d\uff0c\u53ef\u4ee5\u8fdb\u884c\u6269\u5c55\uff0c\u540c\u65f6\u975e\u5e38\u9ad8\u6548\u5730\u5229\u7528\u786c\u4ef6\u3002\u6211\u4e2a\u4eba\u4e5f\u81f4\u529b\u4e8e\u8ba9\u5b83\u66f4\u52a0\u9ad8\u6548\u5730\u5229\u7528\u786c\u4ef6\u3002\u6240\u4ee5\uff0c\u8fd9\u5c31\u662f\u6210\u529f\u7684\u79d8\u8bc0 \u2014\u2014 \u4f5c\u4e00\u4e2a\u65e2\u901a\u7528\u53c8\u80fd\u5f88\u597d\u5730\u8fdb\u884c\u6269\u5c55\u7684\u67b6\u6784\u3002\u5982\u679c\u4f60\u662f\u7814\u7a76 NLP \u7684\uff0c\u4e5f\u8bb8\u4f60\u4f1a\u8003\u8651\u52a0\u5165\u4e00\u4e9b\u5f52\u7eb3\u504f\u7f6e\u6765\u589e\u5f3a\u6a21\u578b\u3002\u5c31\u6211\u4e2a\u4eba\u800c\u8a00\uff0c\u6211\u8ba4\u4e3a Transformer \u662f\u4e00\u4e2a\u975e\u5e38\u901a\u7528\u7684\u67b6\u6784\uff0c\u53ef\u6269\u5c55\u6027\u5f88\u597d\uff0c\u800c\u4e14\u5bf9\u786c\u4ef6\u5f88\u53cb\u597d\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u662f\u7684\uff0c\u662f\u7684\u3002\u73b0\u5728\u56de\u60f3\u8d77\u6765\uff0c\u8fd9\u4e00\u5207\u4f3c\u4e4e\u90fd\u663e\u5f97\u5f88\u660e\u663e\u3002\u73b0\u5728\uff0c\u5728\u7814\u7a76\u5b83\u7684\u66ff\u4ee3\u65b9\u6848\u65f6\uff0c\u4e00\u4e2a\u6709\u8da3\u7684\u7ef4\u5ea6\u662f\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002Michael\uff0c\u4f60\u600e\u4e48\u770b\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Michael Poli<\/strong>\uff1a\u662f\u7684\uff0c\u6211\u6709\u51e0\u70b9\u60f3\u8bf4\u3002\u9996\u5148\uff0c\u76ee\u524d\u4ecd\u7136\u6709\u5f88\u591a\u4f18\u79c0\u7684\u7814\u7a76\u8bd5\u56fe\u4ece\u7b2c\u4e00\u6027\u539f\u7406\u51fa\u53d1\u89e3\u91ca Transformer\uff0c\u4e3a\u4ec0\u4e48\u5b83\u80fd\u5b66\u4e60\u8fd9\u4e9b\u6709\u8da3\u7684\u56de\u8def\uff1f\u4eba\u4eec\u4f1a\u62c6\u89e3\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u6bd4\u5982\u4e0d\u540c\u7684 transformer \u4e2d\u7684 head \u7ec4\u5408\u7b49\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6709\u4e00\u4e9b\u5de5\u4f5c\u4f1a\u5c06 Transformer \u4f5c\u4e3a\u4e00\u79cd\u7f16\u7801\u7684\u7f16\u7a0b\u8bed\u8a00\u6765\u7406\u89e3\u3002\u4f46\u6211\u8ba4\u4e3a\uff0c\u6b63\u5982 Trey \u63d0\u5230\u7684\uff0cTransformer \u4e2d\u6709\u4e00\u4e9b\u975e\u5e38\u6709\u8da3\u7684\u8bbe\u8ba1\u9009\u62e9\u3002\u6ce8\u610f\u529b\u548c MLP \u7684\u4ea4\u7ec7\u4f7f\u7528\u662f\u76f8\u5f53\u91cd\u8981\u7684\u3002\u800c\u4e14\uff0cTransformer \u4e00\u5f00\u59cb\u4e4b\u6240\u4ee5\u6210\u529f\uff0c\u662f\u56e0\u4e3a\u5b83\u91c7\u7528\u4e86\u4e00\u4e9b\u5df2\u7ecf\u4e3a RNN \u548c\u5176\u4ed6\u4f20\u7edf NLP \u6a21\u578b\u6240\u5f00\u53d1\u7684\u6280\u672f\uff0c\u6bd4\u5982\u4f7f\u7528\u95e8\u63a7\u673a\u5236\u6765\u8c03\u8282\u6a21\u578b\u5438\u6536\u54ea\u4e9b\u4fe1\u606f\uff0c\u4ee5\u53ca\u51b3\u5b9a\u67d0\u4e9b\u5185\u5bb9\u5728\u8fd9\u79cd\u5e76\u884c\u5f62\u5f0f\u4e2d\u88ab\u5fd8\u8bb0\u7684\u901f\u5ea6\u3002\u8fd9\u5c31\u50cf\u662f\u4e00\u4e9b\u53ef\u4ee5\u5728 GPU \u4e0a\u8fdb\u884c\u4f18\u5316\u7684\u5b9d\u77f3\uff0c\u867d\u7136\u4e0d\u662f\u5f88\u5bb9\u6613\uff0c\u4f46\u786e\u5b9e\u53ef\u4ee5\u4f18\u5316\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>\u6ce8\u610f\u529b\u8ba1\u7b97\u91cf\u5448\u5e73\u65b9\u7ea7\u589e\u957f<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u662f\u7684\uff0c\u8fd9\u4e9b\u90fd\u5f88\u68d2\u3002\u6211\u60f3\u8bf4\u7684\u6bd4\u8f83\u5177\u4f53\u7684\u4e00\u70b9\u662f\uff0c\u6ce8\u610f\u529b\u673a\u5236\u6700\u7ec8\u5448\u73b0\u51fa\u4e00\u79cd\u968f\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u6210\u4e8c\u6b21\u65b9\u589e\u957f\u7684\u8ba1\u7b97\u6210\u672c\u3002\u5047\u8bbe\u4f60\u6709\u4e00\u4e2a\u957f\u5ea6\u4e3a L \u7684\u8f93\u5165\u5e8f\u5217\uff0c\u5e76\u4e14\u4f60\u60f3\u8f93\u51fa\u4e00\u4e2a\u957f\u5ea6\u4e5f\u4e3a L \u7684\u5e8f\u5217\u3002\u5982\u679c\u4f60\u6df1\u5165\u5230\u6570\u5b66\u7ec6\u8282\u4e2d\u53bb\uff0c\u89c2\u5bdf\u5927\u591a\u6570\u5e93\u5728\u63a8\u7406\u65f6\u53d1\u751f\u7684\u60c5\u51b5\uff0c\u4f60\u4f1a\u53d1\u73b0\u6709\u8fd9\u6837\u4e00\u4e2a\u4e0a\u4e09\u89d2\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u5176\u4e2d\uff0c\u4f60\u53ea\u80fd\u8003\u8651\u6587\u672c\u7684\u8fc7\u53bb\u90e8\u5206\u3002\u968f\u7740\u5904\u7406\u7684\u8fdb\u884c\uff0c\u4f60\u4f1a\u53d1\u73b0\u5b83\u5f62\u6210\u4e86\u4e00\u4e2a L \u5e73\u65b9\u7684\u5173\u7cfb\uff0c\u5176\u4e2d\u7b2c\u4e00\u4e2a token \u53ea\u8003\u8651\u4e00\u4e2a\u5143\u7d20\uff0c\u7136\u540e\u6bcf\u4e2a\u540e\u7eed\u7684 token \u9700\u8981\u8003\u8651\u7684\u8fc7\u53bb token \u9010\u6e10\u589e\u591a\u3002\u521a\u624d\u6211\u4eec\u5df2\u7ecf\u8ba8\u8bba\u4e86 RNN\uff0c\u4ee5\u53ca\u4e00\u4e9b\u975e\u6ce8\u610f\u529b\u7684\u65b9\u6cd5\u662f\u5982\u4f55\u505a\u5230\u4e0d\u53bb\u67e5\u770b\u5e8f\u5217\u4e2d\u6240\u6709\u6587\u672c\u5386\u53f2\u7684\u3002\u5f53\u4f60\u7ed9\u804a\u5929\u673a\u5668\u4eba GPT \u5199\u4e00\u4e2a\u957f\u957f\u7684\u63d0\u793a\u65f6\uff0c\u4f60\u771f\u7684\u5e0c\u671b\u6240\u6709\u90a3\u4e9b\u4fe1\u606f\u90fd\u88ab\u7f16\u7801\u8fdb\u53bb\u5417\uff1f\u9664\u4e86\u8fd9\u79cd\u5bc6\u96c6\u7684\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u6211\u4eec\u8fd8\u6709\u4ec0\u4e48\u5176\u4ed6\u7684\u9009\u62e9\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u53ef\u4ee5\u8ffd\u6eaf\u5230 80 \u5e74\u4ee3\uff0c\u4e5f\u8bb8\u5176\u4e2d\u4e00\u4e9b\u66f4\u8457\u540d\u7684\u662f\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc\uff08LSTM\uff09\u3001\u95e8\u63a7\u5faa\u73af\u5355\u5143\uff08GRU\uff09\u3002\u5927\u7ea6\u5728 2012 \u5e74\u5230 2016 \u5e74\u5de6\u53f3\uff0c\u5b83\u4eec\u5728\u7ffb\u8bd1\u3001\u8bed\u97f3\u8bc6\u522b\u7b49\u65b9\u9762\u975e\u5e38\u6d41\u884c\uff0c\u5f53\u65f6\u5b83\u4eec\u662f NLP \u9886\u57df\u7684 SOTA \u6280\u672f\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5b83\u4eec\u4ee5\u4e00\u79cd\u5e8f\u5217\u65b9\u5f0f\u5904\u7406\u6587\u672c\uff1a\u9010\u4e2a\u89c2\u5bdf token\uff0c\u7136\u540e\u6539\u53d8\u9690\u85cf\u72b6\u6001\uff0c\u6bcf\u6b21\u770b\u5230\u4e00\u4e2a\u65b0 token \u65f6\u90fd\u4f1a\u66f4\u65b0\u9690\u85cf\u72b6\u6001\u3002\u6211\u8ba4\u4e3a\u4ece\u67d0\u79cd\u610f\u4e49\u4e0a\u6765\u8bf4\uff0c\u8fd9\u6a21\u4eff\u4e86\u4eba\u8111\u5904\u7406\u4fe1\u606f\u7684\u65b9\u5f0f\uff0c\u5c31\u50cf\u4f60\u9605\u8bfb\u4e00\u53e5\u8bdd\u6216\u4e00\u6bb5\u8bdd\uff0c\u5c31\u50cf\u4f60\u5728\u5927\u8111\u4e2d\u5b58\u50a8\u4e00\u4e9b\u4fe1\u606f\u3002\u5f53\u4f60\u8bfb\u5b8c\u4e00\u4e2a\u6587\u6863\u65f6\uff0c\u4f60\u53ef\u80fd\u80fd\u591f\u56de\u7b54\u5173\u4e8e\u90a3\u4e2a\u6587\u6863\u7684\u95ee\u9898\uff0c\u800c\u65e0\u9700\u518d\u6b21\u53c2\u8003\u8be5\u6587\u6863\u3002\u6240\u4ee5\uff0cRNN \u5c31\u662f\u8fd9\u6837\u5de5\u4f5c\u7684\u3002\u5b83\u4eec\u5904\u7406\u6587\u672c\uff0c\u7136\u540e\u6539\u53d8\u9690\u85cf\u72b6\u6001\uff0c\u9690\u85cf\u72b6\u6001\u662f\u53ef\u4ee5\u7528\u6765\u751f\u6210\u65b0 token \u6216\u5bf9\u6587\u6863\u8fdb\u884c\u5206\u7c7b\u7684\u8868\u793a\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5927\u7ea6\u5728 2016 \u5e74\u5de6\u53f3\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u6548\u679c\u5f88\u597d\u3002\u4f46\u662f\uff0c\u5b83\u4eec\u9010\u6e10\u4e0d\u53d7\u9752\u7750\uff0c\u4ece\u5b9e\u9a8c\u7ed3\u679c\u6765\u770b\uff0c\u5b83\u4eec\u7684\u8868\u73b0\u5e76\u4e0d\u50cf Transformer \u90a3\u4e48\u597d\u3002\u6b63\u5982\u4f60\u63d0\u5230\u7684\uff0c\u7531\u4e8e Transformer \u5b58\u5728\u8fd9\u79cd\u4e8c\u6b21\u65b9\u6269\u5c55\u7279\u6027\uff0c\u6bcf\u4e2a token \u90fd\u4f1a\u4e0e\u4e4b\u524d\u7684\u5404\u4e2a token \u8fdb\u884c\u6bd4\u8f83\uff0c\u8fd9\u4e3a\u4fe1\u606f\u4f20\u64ad\u63d0\u4f9b\u4e86\u4e00\u79cd\u975e\u5e38\u7b80\u5355\u7684\u65b9\u5f0f\u3002\u6211\u8ba4\u4e3a\u8fd9\u662f Transformer \u548c\u6ce8\u610f\u529b\u673a\u5236\u8868\u73b0\u51fa\u8272\u7684\u90e8\u5206\u539f\u56e0\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4f46\u6700\u8fd1\uff0c\u4e00\u4e9b\u65b0\u7684 RNN \u67b6\u6784\u4f3c\u4e4e\u8868\u73b0\u5f97\u76f8\u5f53\u4e0d\u9519\uff0cRWKV \u662f\u5176\u4e2d\u8f83\u65e9\u7684\u4e00\u4e2a\u3002\u6211\u975e\u5e38\u94a6\u4f69\u90a3\u4e2a\u9879\u76ee\uff0c\u90a3\u662f\u4e00\u4f4d\u540d\u53eb Bo Peng \u7684\u7814\u7a76\u8005\u4e3b\u5bfc\u7814\u53d1\u7684\u3002\u5b83\u4f3c\u4e4e\u5728\u4ee5\u4e00\u5df1\u4e4b\u529b\u5bf9\u6297 Transformer \u7edf\u6cbb\u5730\u4f4d\uff0c\u8bc1\u660e\u4e86 RNN \u53ef\u4ee5\u975e\u5e38\u5f3a\u5927\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u662f\u7684\u3002\u6211\u4e4b\u524d\u4e5f\u8bfb\u8fc7\u8fd9\u7bc7\u8bba\u6587\u3002\u5728\u6280\u672f\u5c42\u9762\u4e0a\uff0c\u4ed6\u4eec\u5c1d\u8bd5\u901a\u8fc7\u4e24\u4e2a\u7ebf\u6027 RNN \u6765\u590d\u5236\u7c7b\u4f3c\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u7684\u67e5\u8be2\u952e\u503c\u67e5\u627e\u7684\u4e1c\u897f\uff0c\u672c\u8d28\u4e0a\u662f\u4e3a\u4e86\u6d88\u9664\u50cf\u7279\u5b9a\u6ce8\u610f\u529b\u6269\u5c55\u8fd9\u6837\u7684\u6f5c\u5728\u95ee\u9898\u3002\u8fd9\u4e24\u4e2a RNN \u62e5\u6709\u66f4\u597d\u7684\u957f\u4e0a\u4e0b\u6587\u884c\u4e3a\u548c\u4e0d\u540c\u7684\u5b9e\u73b0\u89c4\u5219\u3002\u4ed6\u4eec\u8fd8\u8bad\u7ec3\u4e86\u53c2\u6570\u91cf\u9ad8\u8fbe 140 \u4ebf\u7684\u6a21\u578b\u3002\u8fd9\u4e5f\u5f15\u51fa\u4e86\u6211\u63a5\u4e0b\u6765\u60f3\u95ee\u7684\u4e00\u4e9b\u95ee\u9898\uff0c\u5305\u62ec Mamba \u548c Striped Hyena\u3002\u6211\u4eec\u53ef\u4ee5\u4e00\u4e2a\u4e00\u4e2a\u6765\u804a\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>Striped Hyena \u662f\u4e2a\u4ec0\u4e48\u6a21\u578b\uff1f<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u6211\u8fdb\u5165\u4e86 Together API\uff0c\u505a\u4e86\u4e00\u6b21 Mistral \u548c Striped Hyena \u7684\u5bf9\u6bd4\u6d4b\u8bd5\u3002\u7ed3\u679c\u663e\u793a\uff0cStriped Hyena \u662f\u4e00\u4e2a\u4e0d\u9519\u7684\u8bed\u8a00\u6a21\u578b\u3002\u5b83\u80fd\u56de\u7b54\u5927\u591a\u6570\u95ee\u9898\uff0c\u6ca1\u6709\u660e\u663e\u7684\u5931\u8d25\u6a21\u5f0f\u3002Michael\uff0c\u4f60\u600e\u4e48\u770b\u5f85\u8fd9\u4e2a\u6a21\u578b\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Michael Poli<\/strong>\uff1a\u9996\u5148\u6211\u60f3\u8bf4\uff0c\u8fd9\u4e9b\u65b0\u65b9\u6cd5\u4e4b\u95f4\u6709\u4e00\u4e2a\u6709\u8da3\u7684\u8054\u7cfb\u3002\u5b58\u5728\u4e00\u79cd\u51f8\u96c6\uff0c\u5b83\u6709\u4e00\u4e2a\u4e2d\u5fc3\u70b9\uff0c\u7ebf\u6027\u6ce8\u610f\u529b\uff08\u5373\u6ca1\u6709 softmax \u7684\u6ce8\u610f\u529b\uff09\u3001\u7ebf\u6027 RNN \u4ee5\u53ca\u57fa\u4e8e\u72b6\u6001\u7684\u6a21\u578b\uff08SSM\uff09\u4e4b\u95f4\u7684\u5173\u8054\u90fd\u5728\u8fd9\u4e2a\u51f8\u96c6\u91cc\u3002\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u8fd9\u79cd\u57fa\u7840\u6a21\u578b\u7684\u6570\u5b66\u8868\u8ff0\u662f\u76f8\u540c\u7684\uff0c\u6211\u8fd9\u91cc\u4e0d\u662f\u6307\u57fa\u7840\u67b6\u6784\uff0c\u800c\u662f\u6307\u57fa\u7840\u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u7136\u540e\u4f60\u53ef\u4ee5\u671d\u4e0d\u540c\u7684\u65b9\u5411\u53d1\u5c55\uff0c\u6bcf\u4e2a\u65b9\u5411\u90fd\u6709\u81ea\u5df1\u7684\u6743\u8861\uff0c\u6bd4\u5982\u7279\u5f81\u6620\u5c04\u65b9\u5411\u3001 kernel \u65b9\u5411\u3002\u6240\u4ee5\u5f53\u4f60\u5206\u89e3\u6216\u53bb\u9664\u6389 softmax \u65f6\uff0c\u4f60\u53ef\u4ee5\u5728\u5904\u7406\u67e5\u8be2\uff08queries\uff09\u548c\u952e\uff08keys\uff09\u65f6\u91c7\u53d6\u4e0d\u540c\u7684\u65b9\u6cd5\u3002\u8fd9\u4e9b\u67e5\u8be2\u548c\u952e\u662f\u6784\u6210\u4f60\u7684\u6ce8\u610f\u529b\u77e9\u9635\u7684\u57fa\u672c\u5b9e\u4f53\u3002\u5728\u53bb\u9664 softmax \u4e4b\u540e\uff0c\u4f60\u53ef\u4ee5\u6784\u5efa\u5176\u4ed6\u7c7b\u4f3c kernel \u7684\u51fd\u6570\uff0c\u6216\u8005\u5176\u4ed6\u4f60\u5e0c\u671b\u80fd\u591f\u8fd1\u4f3c\u5b9e\u73b0\u6ce8\u610f\u529b\u673a\u5236\u529f\u80fd\u7684\u51fd\u6570\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4f60\u53ef\u4ee5\u505a\u4e00\u4e9b\u7c7b\u4f3c\u6cf0\u52d2\u8fd1\u4f3c\uff08Taylor approximation\uff09\u6216\u6cf0\u52d2\u5c55\u5f00\uff08Taylor expansion\uff09\u8fd9\u6837\u7684\u4e8b\u60c5\u3002\u4f60\u4f1a\u5f97\u5230\u4e00\u4e2a\u7565\u663e\u4e0d\u540c\u7684\u89c6\u89d2\uff0c\u4f46\u4f60\u5f97\u5230\u7684\u4e1c\u897f\u53c8\u975e\u5e38\u76f8\u4f3c\u3002\u4f60\u53ef\u4ee5\u8f6c\u5411 Time variance\u3002\u8fd9\u610f\u5473\u7740\u4f60\u4fee\u6539 RNN\uff0c\u4f7f\u5176\u8ba1\u7b97\u66f4\u591a\u5730\u4f9d\u8d56\u4e8e\u8f93\u5165\u5e8f\u5217\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728\u7ebf\u6027 RNN \u4e2d\u7684\u8ba1\u7b97\u662f\u7531\u8f93\u5165\u5e8f\u5217\u6240\u51b3\u5b9a\u7684\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528\u7c7b\u4f3c\u95e8\u63a7\uff08gates\uff09\u7684\u4e1c\u897f\uff0c\u6211\u4eec\u5df2\u7ecf\u770b\u5230\u4e86\u5f88\u591a\u5de5\u4f5c\uff0c\u4f8b\u5982\uff0c\u7528\u989d\u5916\u7684\u95e8\u63a7\u6765\u66f4\u65b0\u5185\u90e8 tension\uff0c\u8ba9\u4f60\u66f4\u597d\u5730\u5229\u7528\u4f60\u7684\u56fa\u5b9a\u72b6\u6001\u7ef4\u5ea6\u3002\u7b2c\u4e09\u4e2a\u65b9\u5411 \u2014\u2014 \u81f3\u5c11\u5728\u6211\u770b\u6765 \u2014\u2014 \u662f\u4f7f\u7528\u5377\u79ef\u5f62\u5f0f\u7684\uff0c\u66f4\u591a\u5730\u4f7f\u7528\u5176\u4ed6\u7c7b\u578b\u7684\u7ebf\u6027\u8fd0\u7b97\u7b26\uff0c\u8fd9\u4e9b\u8fd0\u7b97\u7b26\u4ecd\u7136\u662f\u53ef\u7ed3\u5408\u7684\uff0c\u4ecd\u7136\u5141\u8bb8\u4f60\u5e76\u884c\u8bad\u7ec3\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6240\u4ee5\u8fd9\u91cc\u7684\u5185\u5bb9\u5305\u62ec\u65f6\u95f4\u4e0d\u53d8\u7cfb\u7edf\u3002\u6211\u53ef\u4ee5\u8be6\u7ec6\u89e3\u91ca\u8fd9\u4e9b\u70b9\uff0c\u4f46\u662f\u5b58\u5728\u4e00\u4e9b\u53ef\u4ee5\u5728\u5377\u79ef\u548c\u5faa\u73af\u4e4b\u95f4\u5207\u6362\u7684\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u8fd8\u914d\u5907\u4e86\u989d\u5916\u7684\u95e8\u63a7\u673a\u5236\u3002\u6211\u53c2\u4e0e\u7684\u4e00\u4e2a\u9879\u76ee\uff0c\u5b83\u8bde\u751f\u4e8e\u6211\u521a\u624d\u63d0\u5230\u7684\u8fd9\u7b2c\u4e09\u7c7b\u67b6\u6784\u3002\u6211\u4eec\u771f\u6b63\u8bd5\u56fe\u505a\u5230\u7684\u662f\uff0c\u6253\u9020\u51fa\u6bcf\u6d6e\u70b9\u8fd0\u7b97\u6027\u80fd\u6700\u4f73\u7684\u67b6\u6784\u3002\u6211\u4eec\u53cd\u590d\u9a8c\u8bc1\u7684\u4e00\u4e2a\u539f\u5219\u662f\uff0c\u4f3c\u4e4e\u5c06\u4e0d\u540c\u5c42\u3001\u4e0d\u540c\u7c7b\u522b\u7684\u6a21\u5757\u751a\u81f3\u662f\u5168\u6ce8\u610f\u529b\u5c42\u7ed3\u5408\u8d77\u6765\uff0c\u5f97\u5230\u7684\u4e1c\u897f\u6bd4\u5355\u72ec\u7684\u7ec4\u4ef6\u8981\u597d\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u6b63\u5728\u8bd5\u56fe\u66f4\u6df1\u5165\u5730\u7406\u89e3\u8fd9\u4e9b\u6a21\u578b\u7684\u7ec4\u5408\u65b9\u9762\u3002\u8fd9\u79cd\u7406\u89e3\u6709\u52a9\u4e8e\u6211\u4eec\u521b\u9020\u6bcf\u6d6e\u70b9\u8fd0\u7b97\u6027\u80fd\u66f4\u4f18\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u4f7f\u7528\u8fd9\u79cd\u6a21\u578b\uff0c\u6211\u4eec\u8fd0\u884c\u4e86\u4e00\u6574\u5957\u7684\u89c4\u6a21\u6269\u5c55\u6cd5\u5219\u3002\u6df7\u5408\u5316\u4e5f\u7ed9\u4e86\u6211\u4eec\u4e00\u4e9b\u4f18\u52bf\uff0c\u56e0\u4e3a\u6211\u4eec\u60f3\u8981\u7684\u662f\u4e00\u79cd\u53ef\u4ee5\u5373\u5f00\u5373\u7528\u7684\u4e1c\u897f\uff0c\u5b83\u4f7f\u5f97\u8fc7\u7a0b\u53d8\u5f97\u66f4\u52a0\u7b80\u5355\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5728\u4e3a\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u8fdb\u884c\u5fae\u8c03\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u5e94\u7528\u4e00\u4e9b\u4e3a Transformers \u5f00\u53d1\u7684\u6280\u672f\uff0c\u800c\u4e14\u6709\u70b9\u51fa\u4eba\u610f\u6599\u7684\u662f\uff0c\u8fd9\u4e9b\u6280\u672f\u5bf9\u4e8e\u6df7\u5408\u4f53\u4e5f\u540c\u6837\u6709\u6548\u3002\u6bd4\u5982\uff0c\u7528\u4e8e\u65cb\u8f6c\u5d4c\u5165\u7684\u7ebf\u6027\u7f29\u653e\u7b49\u7b49\uff0c\u4f60\u53ef\u4ee5\u6df1\u5165\u4e86\u89e3\u7ec6\u8282\u3002\u6240\u4ee5\u8fd9\u4e3b\u8981\u662f\u4e00\u4e2a\u5c1d\u8bd5\u6027\u9879\u76ee\uff0c\u8bd5\u56fe\u5f04\u6e05\u695a\uff0c\u5728\u5f53\u524d\u7684\u73af\u5883\u4e0b\uff0c\u6211\u4eec\u6700\u597d\u80fd\u505a\u5230\u4ec0\u4e48\u7a0b\u5ea6\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>\u4ec0\u4e48\u662f Mamba<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1aStriped Hyena \u4f7f\u7528\u4e86\u4e00\u7ec4\u65b0\u7684\u6a21\u578b\u79fb\u690d\u6280\u672f\uff08model grafting\uff09\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u8bad\u7ec3\u671f\u95f4\u66f4\u6539\u6a21\u578b\u67b6\u6784\uff0c\u5bf9\u6211\u6765\u8bf4\uff0c\u611f\u89c9\u597d\u50cf\u6709\u5f88\u591a\u4e8b\u60c5\u6b63\u5728\u53d1\u751f\uff0c\u6bd4\u5982\u4e00\u4e9b\u4f60\u53ef\u80fd\u4e0d\u80fd\u8c08\u8bba\u592a\u591a\u7684\u5185\u5bb9\u5982\u6570\u636e\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6240\u4ee5\uff0c\u6211\u4e0d\u8ba4\u4e3a\u6240\u6709\u6570\u636e\u90fd\u5f97\u5230\u4e86\u5f88\u597d\u7684\u89e3\u91ca\uff0c\u6bd4\u5982\u8f83\u957f\u7684\u4e0a\u4e0b\u6587\u6570\u636e\u662f\u4ec0\u4e48\uff0c\u4f60\u80fd\u4ece\u6a21\u578b\u7684\u89d2\u5ea6\u89e3\u91ca\u4e00\u4e0b\u5417\uff1f\u6211\u8ba4\u4e3a\u5373\u4f7f\u53ea\u662f\u603b\u7ed3\u4e00\u4e0b\uff0c\u5bf9\u4eba\u4eec\u6765\u8bf4\u4e5f\u662f\u5f88\u597d\u7684\u4f53\u9a8c\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8be5\u9886\u57df\u6709\u5f88\u591a\u5f88\u9177\u7684\u5de5\u4f5c\uff0c\u56e0\u6b64\uff0cAI \u9886\u57df\u6709\u5f88\u591a\u65b0\u7684\u9879\u76ee\u5728\u5c55\u5f00\uff0c\u4f8b\u5982\uff0c\u6709\u4e9b\u4eba\u8bd5\u56fe\u5c06 Lama \u6a21\u578b\u62c6\u89e3\u5e76\u7ee7\u7eed\u8bad\u7ec3\u3002\u4e8b\u5b9e\u4e0a\uff0c\u8fd9\u79cd\u505a\u6cd5\u6709\u70b9\u72c2\u91ce\uff0c\u4eba\u4eec\u5c1d\u8bd5\u91c7\u7528\u5f3a\u5927\u7684\u6a21\u578b\u5e76\u8bd5\u56fe\u8ba9\u5b83\u4eec\u53d8\u5f97\u66f4\u5c0f\uff0c\u540c\u65f6\u8fd8\u60f3\u83b7\u5f97\u5ab2\u7f8e\u5927\u6a21\u578b\u7684\u6027\u80fd\u4f18\u52bf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u867d\u7136\u8fd9\u662f\u4e2a\u9898\u5916\u8bdd\uff0c\u4f46\u6211\u6ca1\u6709\u9884\u6599\u5230\u7684\u662f\u5f53\u4f60\u5173\u6ce8\u793e\u4ea4\u5a92\u4f53\u7684\u65f6\u5019\uff0c\u4f60\u4f1a\u770b\u5230\u4eba\u4eec\u4f1a\u8bf4\uff0c\u54e6\uff0c\u6700\u7ec8\u8fd8\u662f\u72b6\u6001\u975e\u6ce8\u610f\u529b\u6a21\u578b\u8d62\u4e86\u3002\u5728\u6211\u770b\u6765\uff0c\u8fd9\u79cd\u8bf4\u6cd5\u63a9\u76d6\u4e86\u5f88\u591a\u6709\u8da3\u7684\u7ec6\u8282\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u597d\u5427\uff0c\u8ba9\u6211\u4eec\u56de\u5230 Mamba\u3002\u5982\u679c\u6ca1\u8bb0\u9519\u7684\u8bdd\uff0c\u6211\u8ba4\u4e3a Mamba \u5957\u4ef6\u4e2d\u6700\u5927\u7684\u6a21\u578b\u662f 2. 8 \u4ebf\u53c2\u6570\uff0c\u5728 NLP \u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u5305\u62ec GPT J\u3001Pythia \u6a21\u578b\u5957\u4ef6\u7ed9\u51fa\u7684\u57fa\u51c6\u5206\u6570\u90fd\u975e\u5e38\u5f3a\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1aMamba \u662f\u6211\u4e0e Albert Gu \u5408\u4f5c\u5b8c\u6210\u7684\uff0c\u5f53\u65f6\u4ed6\u662f\u65af\u5766\u798f\u5927\u5b66\u7684\u535a\u58eb\u751f\uff0c\u6211\u4eec\u5c31\u662f\u5728\u90a3\u91cc\u8ba4\u8bc6\u7684\uff0c\u4ed6\u73b0\u5728\u662f CMU \u7684\u52a9\u7406\u6559\u6388\u3002\u6240\u4ee5\u8fd9\u662f\u4e00\u6b21\u7f8e\u5999\u7684\u5408\u4f5c\uff0cMamba \u7684\u6210\u529f\u8981\u5f52\u529f\u4e8e\u4ed6\u3002Albert \u4e00\u76f4\u81f4\u529b\u4e8e\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u7684\u7814\u7a76\uff0c\u4ece\u67d0\u79cd\u610f\u4e49\u4e0a\u8bf4\uff0c\u6b63\u5982\u524d\u9762\u63d0\u5230\u7684\uff0c\u7ebf\u6027\u5f20\u91cf\u3001\u7ebf\u6027 RNN\u3001\u5377\u79ef\u3001\u795e\u7ecf\u7f51\u7edc\u7b49\u9886\u57df\u4ed6\u90fd\u6709\u6d89\u53ca\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5728\u6211\u8fc7\u53bb\u53c2\u4e0e\u7684\u51e0\u4e2a\u9879\u76ee\u4e2d\uff0c\u6211\u8fd8\u81f4\u529b\u4e8e\u7a7a\u95f4\u3001\u72b6\u6001\u7a7a\u95f4\u7684\u7814\u7a76\uff0c\u6211\u7814\u7a76\u7684\u89d2\u5ea6\u662f\u5982\u4f55\u4f7f\u72b6\u6001\u7a7a\u95f4\u66f4\u52a0\u786c\u4ef6\u9ad8\u6548\uff0c\u5e76\u63d0\u9ad8\u5176\u6027\u80fd\u3002\u6240\u4ee5\u548c Albert Gu \u4e00\u8d77\u5de5\u4f5c\u771f\u662f\u592a\u68d2\u4e86\u3002\u6211\u8ba4\u4e3a\u53c2\u4e0e Mamba \u7684\u7814\u7a76\u8fc7\u7a0b\u66f4\u591a\u7684\u662f\u4e00\u4e2a\u6982\u5ff5\u8bc1\u660e\uff0c\u72b6\u6001\u7a7a\u95f4\u5b9e\u9645\u4e0a\u53ef\u4ee5\u50cf NLP \u9886\u57df\u7684 transforme \u4e00\u6837\u597d\u5417\uff1f\u56e0\u6b64\uff0c\u624d\u6709\u4e86 Mamba\uff0c\u8fd9\u9879\u7814\u7a76\u8868\u660e\u72b6\u6001\u7a7a\u95f4\u5728\u97f3\u9891\u65b9\u9762\u53ef\u80fd\u4f1a\u66f4\u597d\u3002\u4f46\u662f\uff0c\u5bf9\u4e8e\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u6765\u8bf4\uff0c\u8bed\u8a00\u4e00\u76f4\u662f\u6700\u96be\u83b7\u5f97\u3001\u6700\u96be\u505a\u597d\u7684\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u800c\u4e14\uff0c\u8bed\u8a00\u4e5f\u662f\u4eba\u4eec\u73b0\u5728\u6700\u5173\u5fc3\u7684\u4e8b\u60c5\uff0c\u6240\u4ee5\u6211\u505a\u7684\u66f4\u591a\u7684\u662f\u4e00\u4e2a\u6982\u5ff5\u9a8c\u8bc1\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u6211\u4eec\u60f3\u8868\u660e\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u4e5f\u53ef\u4ee5\u5177\u6709\u7ade\u4e89\u529b\uff0c\u751a\u81f3\u53ef\u4ee5\u548c Transformer \u4e00\u8f83\u9ad8\u4e0b\u3002\u6211\u4eec\u5b9e\u9a8c\u4e2d\u9a8c\u8bc1\u7684 token \u6570\u4ece 3B \u5230 300 B \u4e0d\u7b49\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6240\u4ee5\u4ece\u7edd\u5bf9\u610f\u4e49\u4e0a\u8bb2\uff0c\u8fd9\u4e9b\u90fd\u4e0d\u662f\u975e\u5e38\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u8fd9\u4e9b\u8fd8\u4e0d\u662f\u6211\u4eec\u771f\u6b63\u60f3\u8981\u7684\u6a21\u578b\u3002\u6211\u8ba4\u4e3a\u6211\u4eec\u6b63\u5728\u505a\u7684\u66f4\u591a\u7684\u662f\u8fdb\u884c\u5b66\u672f\u6bd4\u8f83\u3002\u6bd4\u5982\u8bad\u7ec3\u76f8\u540c\u6570\u91cf\u7684 token\uff0c\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u53ef\u80fd\u6bd4 transformer \u7a0d\u5fae\u597d\u4e00\u4e9b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u4ef6\u4e8b\u5bf9\u6211\u4eec\u6765\u8bf4\u5c24\u5176\u4ee4\u4eba\u5174\u594b\uff0c\u6211\u8ba4\u4e3a Albert \u5df2\u7ecf\u63a8\u52a8\u8fd9\u4ef6\u4e8b\u6709\u4e00\u6bb5\u65f6\u95f4\u4e86\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u7ed3\u679c\u662f\uff0c\u6211\u4eec\u7684\u7814\u7a76\u5728\u63a8\u7406\u4e0a\u53ef\u80fd\u4f1a\u66f4\u5feb\uff0c\u4e5f\u8bb8\u6211\u4eec\u4f1a\u6709\u4e0d\u540c\u7684\u65b9\u5f0f\u6765\u7406\u89e3\u4e0a\u4e0b\u6587\u5b66\u4e60\u662f\u5982\u4f55\u53d1\u751f\u7684\u3002\u6211\u5f88\u671f\u5f85\u672a\u6765\u7684\u5de5\u4f5c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>Mamba \u786c\u4ef6\u4f18\u5316<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u60a8\u80fd\u5426\u8c08\u4e00\u8c08\u5b9e\u73b0\u8fd9\u4e9b\u65b0 CUDA \u5185\u6838\u5b9e\u9645\u4e0a\u9700\u8981\u505a\u4ec0\u4e48\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u5173\u4e8e\u72b6\u6001\u7a7a\u95f4\u7684\u7814\u7a76\uff0c\u5b83\u5728\u67d0\u79cd\u610f\u4e49\u4e0a\u662f\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u3002\u72b6\u6001\u5927\u5c0f\u662f\u4f60\u5728\u904d\u5386\u6216\u5904\u7406\u5e8f\u5217\u65f6\uff0c\u7528\u6765\u5b58\u50a8\u4fe1\u606f\u7684\u7f13\u51b2\u533a\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4ece\u67d0\u79cd\u610f\u4e49\u4e0a\u8bb2\uff0cTransformer \u4e5f\u53ef\u4ee5\u8fd9\u6837\u7406\u89e3\uff0c\u5b83\u4fdd\u5b58\u7684\u6574\u4e2a\u5386\u53f2\u8bb0\u5f55\u901a\u5e38\u79f0\u4e3a KV \u7f13\u5b58\uff0cTransformer \u4fdd\u7559\u5386\u53f2\u5e76\u4e0d\u65ad\u5f15\u7528\u5b83\u3002\u5bf9\u4e8e RNN\uff0c\u5b83\u4eec\u5177\u6709\u56fa\u5b9a\u5927\u5c0f\u7684\u72b6\u6001\uff1b\u5bf9\u4e8e transformer\uff0c\u4f60\u53ef\u4ee5\u8ba4\u4e3a\u72b6\u6001\u5927\u5c0f\u6b63\u5728\u589e\u52a0\u3002\u800c\u4e14\uff0c\u6211\u4eec\u7684\u76f4\u89c9\u662f\uff0c\u72b6\u6001\u89c4\u6a21\u8d8a\u5927\uff0c\u6a21\u578b\u5c31\u8868\u73b0\u8d8a\u597d\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6240\u4ee5\u57fa\u672c\u4e0a\uff0c\u4f60\u9700\u8981\u6709\u66f4\u591a\u7684\u7a7a\u95f4\u6765\u5b58\u50a8\u4f60\u9700\u8981\u8bb0\u4f4f\u7684\u4e1c\u897f\u3002\u56e0\u6b64\uff0c\u4e4b\u524d\u7684\u6a21\u578b\uff08\u4f8b\u5982 S4 \u7b49\uff09\u5177\u6709\u76f8\u5f53\u5927\u7684\u9690\u5f0f\u72b6\u6001\u5927\u5c0f\uff0c\u5b83\u4eec\u4f7f\u7528\u5377\u79ef\u89c6\u56fe\u6765\u907f\u514d\u5c06\u72b6\u6001\u5177\u4f53\u5316\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u5e0c\u671b\u5728\u5faa\u73af\u4e2d\u52a0\u5165\u66f4\u591a\u7684\u8f93\u5165\u4f9d\u8d56\u6027\uff0c\u7136\u800c\uff0c\u8fd9\u6837\u505a\u963b\u6b62\u4e86\u6211\u4eec\u4f7f\u7528\u53ef\u4ee5\u63d0\u9ad8\u6548\u7387\u7684\u5377\u79ef\u89c6\u56fe\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u5fc5\u987b\u627e\u5230\u4e00\u79cd\u4e0d\u540c\u7684\u65b9\u6cd5\u6765\u63d0\u9ad8\u6548\u7387\uff0c\u6240\u4ee5\u6211\u4eec\u628a\u91cd\u70b9\u4e13\u6ce8\u4e8e\u5728 GPU \u4e0a\u63d0\u9ad8\u6548\u7387\u3002\u6211\u4eec\u7684\u60f3\u6cd5\u662f\uff0c\u6211\u4eec\u8981\u6709\u4e00\u4e2a\u5927\u7684\u72b6\u6001\u5927\u5c0f\uff0c\u4f46\u6211\u4eec\u4e0d\u9700\u8981\u7528\u5230\u5b9e\u9645\u7684 GPU \u5185\u5b58\uff0c\u6bd4\u5982 HBM\uff0c\u6211\u4eec\u53ef\u4ee5\u628a\u5927\u7684\u72b6\u6001\u4fdd\u5b58\u5728\u4e00\u4e2a\u66f4\u5feb\u7684\u5185\u5b58\u4e2d\uff0c\u79f0\u4e4b\u4e3a SRAM\uff0c\u4f60\u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u4e00\u4e2a\u7f13\u5b58\u3002\u5982\u679c\u4f60\u66f4\u719f\u6089 CPU\uff0c\u8fd9\u901a\u5e38\u662f\u4e00\u4e2a\u7f13\u5b58\u548c RAM\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6240\u4ee5\uff0c\u5982\u679c\u4f60\u6709\u8f83\u5927\u7684\u72b6\u6001\uff0c\u4f60\u53ef\u4ee5\u5c06\u5b83\u4fdd\u5b58\u5728\u7f13\u5b58\u4e2d\uff0c\u8fd9\u6837\u4e00\u6765\u4f60\u4e0d\u4f1a\u906d\u53d7\u592a\u5927\u7684\u635f\u5931\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>2024 \u5e74\u67b6\u6784\u9884\u6d4b<\/h4>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u6211\u73b0\u5728\u5bf9 GPU \u4e0e TPU \u6700\u6709\u529b\u7684\u6d1e\u5bdf\u662f\uff0cMoE \u5728 TPU \u4e2d\u6548\u679c\u4e0d\u4f73\uff0c\u56e0\u4e3a\u4f60\u5fc5\u987b\u5728\u57fa\u672c\u5c42\u4e0a\u52a0\u5165\u4e00\u4e9b MoE\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e2d\uff0c\u524d\u9988\u5c42\u53ef\u80fd\u6700\u7ec8\u5206\u5e03\u5728\u4e0d\u540c\u7684 TPU \u8282\u70b9\u4e0a\uff0c\u800c TPU \u4e4b\u95f4\u662f\u901a\u8fc7\u90bb\u8fd1\u8282\u70b9\u8fdb\u884c\u901a\u4fe1\u7684\u3002\u56e0\u6b64\uff0c\u4e0e GPU \u76f8\u6bd4\uff0cTPU \u5728\u8fd9\u65b9\u9762\u4f1a\u53d7\u5230\u8f83\u5927\u7684\u5f71\u54cd\u30022024 \u5e74\u8fd9\u4e2a\u9886\u57df\u4f1a\u53d1\u751f\u4ec0\u4e48\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u6211\u8ba4\u4e3a Transform \u4ecd\u7136\u662f\u4e00\u4e2a\u975e\u5e38\u5f3a\u5927\u7684\u67b6\u6784\uff0c\u73b0\u5728\u53ef\u4ee5\u6269\u5c55\u5230\u4e00\u4e07\u4ebf\u7ea7\u522b\u7684\u53c2\u6570\uff0c\u4eba\u4eec\u5f80\u5f80\u60f3\u8981\u6027\u80fd\u6700\u597d\u7684\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728\u786c\u4ef6\u4e0a\u8fd0\u884c\u6548\u7387\u6700\u9ad8\uff0c\u5e76\u4e14\u5728\u8f6f\u4ef6\u65b9\u9762\u62e5\u6709\u6700\u591a\u7684\u652f\u6301\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u6700\u8fd1\u6709\u4e9b\u65b0\u60f3\u6cd5\uff0c\u6bd4\u5982\u72b6\u6001\u7a7a\u95f4\u3002\u6211\u4eec\u5df2\u7ecf\u770b\u5230\uff0c\u6b63\u5982 Michael \u63d0\u5230\u7684\uff0c\u6df7\u5408\u8fd9\u4e9b\u7ec4\u4ef6\u4f3c\u4e4e\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\uff0c\u6211\u8ba4\u4e3a\u5728 7B \u5927\u5c0f\u7684\u6a21\u578b\u4e0a\u5f97\u5230\u4e86\u9a8c\u8bc1\uff0c\u4e5f\u8bb8\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u53ef\u4ee5\u5728\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u4e0a\u8d77\u4f5c\u7528\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5f53\u524d\u5927\u591a\u6570\u4eba\u5173\u6ce8\u7684\u662f\u57fa\u4e8e Lime \u67b6\u6784\u7684\u6570\u636e\u548c\u57fa\u7840\u8bbe\u65bd\u5efa\u8bbe\uff0c\u5c3d\u7ba1\u73b0\u6709\u7684 Transformer \u67b6\u6784\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u4ecd\u7136\u975e\u5e38\u5f3a\u5927\u4e14\u5f97\u5230\u5e7f\u6cdb\u652f\u6301\uff0c\u4f46\u4e5f\u5b58\u5728\u4e00\u4e9b\u8fb9\u7f18\u9886\u57df\uff0c\u6bd4\u5982\u957f\u4e0a\u4e0b\u6587\u3001\u97f3\u9891\u3001\u57fa\u56e0\u7ec4\u5b66\u7b49\uff0c\u7814\u7a76\u8fd9\u4e9b\u9886\u57df\u4e2d\u7684\u66ff\u4ee3\u67b6\u6784\u4f1a\u975e\u5e38\u6709\u8da3\u3002\u8fd9\u4e9b\u9886\u57df\u63d0\u51fa\u4e86\u4e00\u4e9b\u6709\u610f\u4e49\u7684\u79d1\u5b66\u95ee\u9898\uff0c\u4f8b\u5982\u6a21\u578b\u662f\u5426\u50cf\u4eba\u7c7b\u4e00\u6837\u7406\u89e3\u6307\u4ee4\u548c\u76f4\u89c9\u3001\u662f\u5426\u80fd\u4e0e\u91cf\u5316\u65b9\u6cd5\u534f\u540c\u5de5\u4f5c\u7b49\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6b64\u5916\uff0c\u5373\u4f7f\u73b0\u5728\u4eba\u4eec\u8fd8\u5728\u4f7f\u7528 Transformer \u67b6\u6784\uff0c\u672a\u6765\u53ef\u80fd\u4f1a\u878d\u5165\u66f4\u591a\u7684\u65b0\u601d\u60f3\u548c\u7ec4\u4ef6\uff0c\u4f8b\u5982\u589e\u52a0\u66f4\u591a\u7684\u5c42\u548c\u6ce8\u610f\u529b\u673a\u5236\u7b49\uff0c\u5c3d\u7ba1\u5b83\u4eec\u53ef\u80fd\u4ecd\u7136\u88ab\u79f0\u4e3a Transformer\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u5c3d\u7ba1\u76ee\u524d\u7684\u4eba\u5de5\u667a\u80fd\u9886\u57df\u503e\u5411\u4e8e\u4fdd\u5b88\uff0c\u4e13\u6ce8\u4e8e\u73b0\u4ee3\u67b6\u6784\uff0c\u4f46\u65b0\u7684\u67b6\u6784\u548c\u601d\u60f3\u6b63\u5728\u9010\u6e10\u6d8c\u73b0\uff0c\u8fd9\u4e9b\u65b0\u9896\u7684\u89c2\u70b9\u548c\u65b9\u6cd5\u53ef\u80fd\u4f1a\u4e3a\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\u5e26\u6765\u65b0\u7684\u52a8\u529b\u548c\u65b9\u5411\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Michael Poli<\/strong>\uff1a\u662f\u7684\uff0c\u6211\u767e\u5206\u4e4b\u767e\u540c\u610f Tri Dao \u7684\u770b\u6cd5\uff0c\u6ce8\u610f\u529b\u673a\u5236\u4f5c\u4e3a\u8ba1\u7b97\u539f\u8bed\u4f9d\u7136\u5f88\u91cd\u8981\u3002\u6ce8\u610f\u529b\u673a\u5236\u4f5c\u4e3a\u4e00\u79cd\u9ad8\u6548\u4fbf\u6377\u7684\u65b9\u5f0f\uff0c\u53ef\u4ee5\u6709\u6548\u589e\u52a0\u5e8f\u5217\u5904\u7406\u5668\u7684\u72b6\u6001\u5bb9\u91cf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u72b6\u6001\u7ef4\u5ea6\u4e0e\u5e8f\u5217\u957f\u5ea6\u4e4b\u95f4\u5b58\u5728\u6743\u8861\u5173\u7cfb\uff0c\u5f53\u6a21\u578b\u5c3a\u5bf8\u53d8\u5927\uff0c\u4e5f\u5c31\u662f\u6a21\u578b\u53d8\u5f97\u66f4\u5bbd\u65f6\uff0c\u5c06\u6709\u6548\u5730\u5f15\u5165\u66f4\u591a\u7684\u72b6\u6001\u548c\u5e8f\u5217\u957f\u5ea6\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u4e00\u4e9b\u8fb9\u9645\u6548\u5e94\u53ef\u80fd\u4f1a\u6d88\u5931\uff0c\u800c\u67d0\u4e9b\u6743\u8861\u5c06\u4f1a\u53d1\u751f\u6539\u53d8\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u90a3\u4e9b\u975e\u5e38\u5927\u7684\u6a21\u578b\uff0c\u5982 14B\u300130B \u8fd9\u7c7b\u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u672a\u6765\uff0c\u67b6\u6784\u8bbe\u8ba1\u5c06\u53d8\u5f97\u66f4\u52a0\u6709\u8da3\u548c\u590d\u6742\uff0c\u5c06\u4f1a\u6709\u66f4\u591a\u7684\u521b\u65b0\u53d1\u751f\u3002\u65e0\u8bba\u662f\u6df7\u5408\u6a21\u578b\u8fd8\u662f\u5f15\u5165\u65b0\u7684\u6a21\u5757\uff0c\u6211\u4eec\u90fd\u5c06\u770b\u5230\u66f4\u591a\u6fc0\u52a8\u4eba\u5fc3\u7684\u521b\u65b0\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>\u5bf9 AI \u66f4\u591a\u7684\u9884\u6d4b<\/h4>\n<p>&nbsp;<\/p>\n<p>Nathan Lambert\uff1a\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Mixture of Experts, MoE\uff09\u548c\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u5728\u8fd1\u671f\u5185\u5448\u73b0\u51fa\u4e00\u79cd\u6d41\u884c\u7684\u8d8b\u52bf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u7136\u800c\uff0c\u5728\u5f00\u6e90\u548c\u5b66\u672f\u754c\uff0c\u76ee\u524d\u8fd8\u6ca1\u6709\u4eba\u771f\u6b63\u5c1d\u8bd5\u5728\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u4e0a\u505a\u65e9\u671f\u7684\u5c1d\u8bd5\u548c\u6539\u8fdb\u3002\u6a21\u578b\u5ac1\u63a5\uff08Model Grafting\uff09\u73b0\u5728\u6b63\u5728\u53d8\u5f97\u66f4\u52a0\u5b9e\u7528\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8ddf\u8fdb\u8fd9\u4e9b\u8fdb\u5c55\u975e\u5e38\u6709\u8da3\uff0c\u5e0c\u671b\u8fd9\u4e9b\u53d1\u5c55\u80fd\u591f\u4e3a\u5b66\u672f\u754c\u548c\u79d1\u5b66\u5bb6\u4eec\u63d0\u4f9b\u66f4\u591a\u5f71\u54cd\u884c\u4e1a\u5bf9\u8bdd\u7684\u65b9\u6cd5\uff0c\u7279\u522b\u662f\u5728\u76ee\u524d\u884c\u4e1a\u66f4\u591a\u5173\u6ce8\u4e8e\u6269\u5927\u6a21\u578b\u89c4\u6a21\u7684\u60c5\u51b5\u4e0b\u3002\u6211\u5efa\u8bae\u5f00\u6e90\u516c\u53f8\u5e94\u8be5\u5728\u4ed6\u4eec\u7684\u8bed\u8a00\u6a21\u578b\u4e2d\u8fdb\u884c\u7279\u5b9a\u7684\u6539\u8fdb\uff0c\u4ee5\u83b7\u5f97\u5546\u4e1a\u4f18\u52bf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4f60\u4eec\u5728\u673a\u5668\u5b66\u4e60\u65b9\u9762\u8fd8\u5173\u6ce8\u5176\u4ed6\u4ec0\u4e48\u5417\uff1f\u5b83\u4e0d\u4e00\u5b9a\u662f\u5173\u4e8e\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u7684\u3002\u660e\u5e74\u6700\u8ba9\u4f60\u4eec\u5174\u594b\u7684\u662f\u4ec0\u4e48\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u6211\u4e2a\u4eba\u8ba4\u4e3a\u6570\u636e\u4ecd\u7136\u662f\u6700\u91cd\u8981\u7684\u56e0\u7d20\u3002\u6211\u4eec\u6b63\u5728\u6df1\u5165\u7814\u7a76\u6570\u636e\u5982\u4f55\u5f71\u54cd\u6a21\u578b\u6027\u80fd\uff0c\u4f8b\u5982\u901a\u8fc7\u4e00\u4e9b\u4e0e\u6a21\u578b\u6027\u80fd\u9ad8\u5ea6\u76f8\u5173\u7684\u5408\u6210\u4efb\u52a1\u6765\u8fdb\u884c\u7814\u7a76\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e00\u76f4\u662f\u6211\u4eec\u8bba\u6587\u548c\u7814\u7a76\u5de5\u4f5c\u4e2d\u7684\u4e3b\u8981\u52a8\u529b\u548c\u793a\u4f8b\u3002\u672a\u6765\u4e00\u6bb5\u65f6\u95f4\u6211\u4eec\u5c06\u4e13\u6ce8\u4e8e\u6570\u636e\u9886\u57df\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u867d\u7136\u6240\u6709\u7684\u67b6\u6784\u5de5\u4f5c\u90fd\u5f88\u6709\u8da3\uff0c\u4f7f\u5176\u5728\u786c\u4ef6\u4e0a\u9ad8\u6548\u8fd0\u884c\u4e5f\u5f88\u6709\u8da3\uff0c\u4f46\u6700\u7ec8\u8fd8\u662f\u5173\u4e8e\u6570\u636e\u7684\u3002\u5982\u679c\u4f60\u4e86\u89e3\u6269\u5c55\u7387\uff08scaling law\uff09\uff0c\u4f60\u5c31\u4f1a\u77e5\u9053\u4e0d\u540c\u7684\u6a21\u578b\u67b6\u6784\u901a\u5e38\u4f1a\u6709\u76f8\u540c\u7684\u659c\u7387\uff0c\u53ea\u662f\u504f\u79fb\u91cf\u4e0d\u540c\uff0c\u4f3c\u4e4e\u552f\u4e00\u80fd\u6539\u53d8\u659c\u7387\u7684\u662f\u6570\u636e\u8d28\u91cf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Michael Poli<\/strong>\uff1a\u662f\u7684\uff0c\u6211\u4eec\u6dfb\u52a0\u4e86\u6570\u636e\u3002\u6570\u636e\u786e\u5b9e\u5f88\u6709\u8da3\uff0c\u6bd4\u5982\u5728\u67b6\u6784\u8bbe\u8ba1\u4e0a\u7684\u5fae\u578b\u5316\uff0c\u627e\u51fa\u5e76\u5206\u89e3\u8bed\u8a00\u5efa\u6a21\u7b49\u4efb\u52a1\u6240\u6d89\u53ca\u7684\u5404\u4e2a\u65b9\u9762\uff0c\u6211\u4eec\u8bd5\u56fe\u5c06\u5b83\u4eec\u6253\u5305\u6210\u53ef\u4ee5\u7528\u6765\u8fed\u4ee3\u7684\u4e1c\u897f\uff0c\u8fd9\u975e\u5e38\u4ee4\u4eba\u5174\u594b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u4e2a\u4eba\u5bf9\u65b0\u5e94\u7528\uff0c\u5c24\u5176\u662f\u57fa\u56e0\u7ec4\u5b66\u5de5\u4f5c\u611f\u5230\u975e\u5e38\u5174\u594b\uff0c\u4f46\u66f4\u591a\u5730\u662f\u4ece\u5de5\u7a0b\u89d2\u5ea6\u51fa\u53d1\uff0c\u6211\u4eec\u6b63\u770b\u5230\u4e00\u4e2a\u8f6c\u53d8\u3002\u76ee\u524d\uff0c\u8bed\u8a00\u4ecd\u662f\u83b7\u5f97\u6700\u591a\u70b9\u51fb\u548c\u6700\u5927\u5174\u8da3\u7684\u9886\u57df\uff0c\u4f46\u6211\u8ba4\u4e3a\u8fd9\u5c06\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u800c\u6539\u53d8\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u662f\u7684\uff0c\u6bcf\u4e2a\u4eba\u90fd\u5728\u8c08\u8bba\u8bed\u8a00\uff0c\u4f46\u6211\u89c9\u5f97\u56fe\u50cf\u3001\u89c6\u9891\u5c06\u4f1a\u662f\u4ea7\u751f\u5de8\u5927\u4ef7\u503c\u7684\u4e1c\u897f\u3002\u6211\u4e0d\u77e5\u9053\u8bed\u8a00\u7684\u4e0a\u9650\u5728\u54ea\u91cc\u3002\u6211\u5f88\u5174\u594b\uff0c\u6211\u5df2\u7ecf\u5f00\u59cb\u5c1d\u8bd5\u8fd9\u4e9b\u4e86\uff0c\u6bd4\u5982\u6211\u4f1a\u62ff\u535a\u5ba2\u7684\u6587\u672c\uff0c\u8ba9\u6a21\u578b\u8f6c\u6362\u6210\u56fe\u7247\uff0c\u7136\u540e\u8f6c\u6362\u6210\u5e26\u6709\u97f3\u9891\u7684\u89c6\u9891\uff0c\u6240\u6709\u8fd9\u4e9b\u90fd\u662f\u7528\u4e00\u4e2a Python \u811a\u672c\u5b8c\u6210\u7684\uff0c\u8fd9\u771f\u7684\u5f88\u5bb9\u6613\u505a\u5230\u3002\u6240\u4ee5\u6211\u540c\u610f\u4f60\u7684\u89c2\u70b9\uff0c\u8d85\u8d8a\u8bed\u8a00\u7684\u4e1c\u897f\u5f88\u6709\u8da3\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u5728\u4f60\u7684\u7ecf\u9a8c\u4e2d\uff0c\u5f53\u4f60\u5c06\u6240\u6709\u8fd9\u4e9b\u4e1c\u897f\u62fc\u51d1\u5728\u4e00\u8d77\u65f6\uff0c\u5b83\u4eec\u786e\u5b9e\u80fd\u591f\u5408\u7406\u5730\u5de5\u4f5c\u5417\uff1f<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Nathan Lambert<\/strong>\uff1a\u5b83\u8fd8\u4e0d\u662f\u90a3\u4e48\u5b8c\u7f8e\uff0cDALL\u30fbE \u751f\u6210\u7684\u56fe\u7247\u6bd4\u8f83\u76f8\u4f3c\uff0c\u4f46\u6211\u7684\u505a\u6cd5\u975e\u5e38\u7b80\u5355\uff0c\u5c31\u662f\u76f4\u63a5\u628a\u6587\u672c\u62ff\u8fc7\u6765\uff0c\u7136\u540e\u7528\u4e00\u4e2a\u7cfb\u7edf\u63d0\u793a\uff0c\u8ba9\u6a21\u578b\u751f\u6210\u5404\u79cd\u5404\u6837\u7684\u56fe\u7247\uff0c\u6211\u8ba4\u4e3a\u6211\u53ef\u4ee5\u505a\u5f97\u66f4\u597d\u3002\u636e\u6211\u77e5\u9053\uff0c\u5927\u6982\u5728\u4e00\u5e74\u5185\uff0c\u5c31\u4f1a\u6709\u4e00\u4e2a\u6587\u672c\u5230\u89c6\u9891\u7684 API\uff0c\u5230\u65f6\u6211\u4f1a\u5207\u6362\u5230 API\uff0c\u90a3\u79cd\u4f53\u9a8c\u4f1a\u5f88\u68d2\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>Tri Dao<\/strong>\uff1a\u662f\u7684\uff0c\u6211\u8ba4\u4e3a\u8fd9\u4e9b\u8fdb\u6b65\u786e\u5b9e\u4ea7\u751f\u4e86\u5927\u91cf\u7684\u7ecf\u6d4e\u4ef7\u503c\uff0c\u6211\u4eec\u5df2\u7ecf\u770b\u5230\u4e86\u8fd9\u4e00\u70b9\u3002\u73b0\u5728\u5f88\u591a\u516c\u53f8\u90fd\u5f00\u59cb\u8f6c\u5411\u4f7f\u7528\u8fd9\u4e9b\u6280\u672f\u3002\u6211\u8ba4\u4e3a\u8fd9\u5c06\u6539\u53d8\u6211\u4eec\u7684\u5de5\u4f5c\u65b9\u5f0f\uff0c\u6b63\u5982\u4f60\u63d0\u5230\u7684\uff0c\u6539\u53d8\u6211\u4eec\u7684\u5de5\u4f5c\u65b9\u5f0f\u548c\u5a31\u4e50\u65b9\u5f0f\u3002\u6240\u4ee5\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u4ee4\u4eba\u5174\u594b\u7684\u672a\u6765\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/www.interconnects.ai\/p\/interviewing-tri-dao-and-michael?cnotallow=5d10d34c97637bebcfeba6470c0f0d9b<\/p>\n<\/div>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_17231\" class=\"pvc_stats total_only  \" data-element-id=\"17231\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u6700\u8fd1\u7684\u4e00\u6b21\u91c7\u8bbf\u4e2d\uff0c\u6765\u81ea\u827e\u4f26\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u6240\uff08 AI2\uff09\u7684\u673a\u5668\u5b66\u4e60\u7814\u7a76\u5458 Nathan Lambert \u4e0e Ma [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_17231\" class=\"pvc_stats total_only  \" data-element-id=\"17231\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":17232,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,80],"tags":[124,113,667],"class_list":["post-17231","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-80","tag-ai","tag-113","tag-667"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/17231","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=17231"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/17231\/revisions"}],"predecessor-version":[{"id":17233,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/17231\/revisions\/17233"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/17232"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=17231"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=17231"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=17231"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}