{"id":35136,"date":"2024-12-19T08:59:50","date_gmt":"2024-12-19T00:59:50","guid":{"rendered":"https:\/\/17aitech.com\/?p=35136"},"modified":"2024-12-19T08:59:50","modified_gmt":"2024-12-19T00:59:50","slug":"%e8%8b%b1%e4%bc%9f%e8%be%be%e7%8e%a9%e8%bd%ac%e5%89%aa%e6%9e%9d%e3%80%81%e8%92%b8%e9%a6%8f%ef%bc%9a%e6%8a%8allama-3-1-8b%e5%8f%82%e6%95%b0%e5%87%8f%e5%8d%8a%ef%bc%8c%e6%80%a7%e8%83%bd%e5%90%8c","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35136","title":{"rendered":"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-16-4\" target=\"_blank\">\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a<\/a><\/p>\n<section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"7\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\u5c0f\u6a21\u578b\u5d1b\u8d77\u4e86\u3002<\/section>\n<\/section>\n<\/blockquote>\n<\/section>\n<p>\u4e0a\u4e2a\u6708\uff0cMeta \u53d1\u5e03\u4e86 Llama 3.1 \u7cfb\u5217\u6a21\u578b\uff0c\u5176\u4e2d\u5305\u62ec Meta \u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u7684 405B \u6a21\u578b\uff0c\u4ee5\u53ca\u4e24\u4e2a\u8f83\u5c0f\u7684\u6a21\u578b\uff0c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\u5206\u522b\u4e3a 700 \u4ebf\u548c 80 \u4ebf\u3002<\/p>\n<p>Llama 3.1 \u88ab\u8ba4\u4e3a\u662f\u5f15\u9886\u4e86\u5f00\u6e90\u65b0\u65f6\u4ee3\u3002\u7136\u800c\uff0c\u65b0\u4e00\u4ee3\u7684\u6a21\u578b\u867d\u7136\u6027\u80fd\u5f3a\u5927\uff0c\u4f46\u90e8\u7f72\u65f6\u4ecd\u9700\u8981\u5927\u91cf\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u4e1a\u754c\u51fa\u73b0\u4e86\u53e6\u4e00\u79cd\u8d8b\u52bf\uff0c\u5373\u5f00\u53d1\u5c0f\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark> (SLM)\uff0c\u8fd9\u79cd\u6a21\u578b\u5728\u8bb8\u591a\u8bed\u8a00\u4efb\u52a1\u4e2d\u8868\u73b0\u8db3\u591f\u51fa\u8272\uff0c\u90e8\u7f72\u8d77\u6765\u4e5f\u975e\u5e38\u4fbf\u5b9c\u3002<\/p>\n<p>\u6700\u8fd1\uff0c\u82f1\u4f1f\u8fbe\u7814\u7a76\u8868\u660e\uff0c\u7ed3\u6784\u5316<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark><mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e0e<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\u76f8\u7ed3\u5408\uff0c\u53ef\u4ee5\u4ece\u521d\u59cb\u8f83\u5927\u7684\u6a21\u578b\u4e2d\u9010\u6b65\u83b7\u5f97\u8f83\u5c0f\u7684<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe\u7075\u5956\u5f97\u4e3b\u3001Meta \u9996\u5e2d AI \u79d1\u5b66\u5bb6 Yann LeCun \u4e5f\u70b9\u8d5e\u8f6c\u5e16\u4e86\u8be5\u7814\u7a76\u3002<\/sup><\/em><\/p>\n<p>\u7ecf\u8fc7<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\uff0c\u82f1\u4f1f\u8fbe\u7814\u7a76\u56e2\u961f\u5c06 Llama 3.1 8B \u63d0\u70bc\u4e3a Llama-3.1-Minitron 4B \u5f00\u6e90\u4e86\u51fa\u6765\u3002\u8fd9\u662f\u82f1\u4f1f\u8fbe\u5728 Llama 3.1 \u5f00\u6e90\u7cfb\u5217\u4e2d\u7684\u7b2c\u4e00\u4e2a\u4f5c\u54c1\u3002<\/p>\n<p>Llama-3.1-Minitron 4B \u7684\u8868\u73b0\u4f18\u4e8e\u7c7b\u4f3c\u5927\u5c0f\u7684\u6700\u5148\u8fdb\u7684\u5f00\u6e90\u6a21\u578b\uff0c\u5305\u62ec Minitron 4B\u3001Phi-2 2.7B\u3001Gemma2 2.6B \u548c Qwen2-1.5B\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-dafd407616fa932976cc2ea77ad921e1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-dafd407616fa932976cc2ea77ad921e1.png\"><\/a><\/section>\n<p>\u8fd9\u9879\u7814\u7a76\u7684\u76f8\u5173\u8bba\u6587\u65e9\u5728\u4e0a\u4e2a\u6708\u5df2\u7ecf\u653e\u51fa\u4e86\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-5e9e93a7f516af40c6f52ccceb76186e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-5e9e93a7f516af40c6f52ccceb76186e.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/www.arxiv.org\/pdf\/2407.14679<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aCompact Language Models via Pruning and Knowledge Distillation<\/p>\n<\/li>\n<\/ul>\n<p><strong><mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f<\/strong><\/p>\n<p><mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4f7f\u6a21\u578b\u53d8\u5f97\u66f4\u5c0f\u3001\u66f4\u7cbe\u7b80\uff0c\u53ef\u4ee5\u901a\u8fc7\u5220\u9664\u5c42\uff08\u6df1\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\uff09\u6216\u5220\u9664<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u548c\u6ce8\u610f\u529b\u5934\u4ee5\u53ca\u5d4c\u5165\u901a\u9053\uff08\u5bbd\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\uff09\u6765\u5b9e\u73b0\u3002<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u901a\u5e38\u4f34\u968f\u7740\u4e00\u5b9a\u7a0b\u5ea6\u7684\u518d\u8bad\u7ec3\uff0c\u4ee5\u6062\u590d<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u3002<\/p>\n<p>\u6a21\u578b\u84b8\u998f\u662f\u4e00\u79cd\u5c06\u77e5\u8bc6\u4ece\u5927\u578b\u590d\u6742\u6a21\u578b\uff08\u901a\u5e38\u79f0\u4e3a\u6559\u5e08\u6a21\u578b\uff09\u8fc1\u79fb\u5230\u8f83\u5c0f\u3001\u8f83\u7b80\u5355\u7684\u5b66\u751f\u6a21\u578b\u7684\u6280\u672f\u3002\u76ee\u6807\u662f\u521b\u5efa\u4e00\u4e2a\u66f4\u9ad8\u6548\u7684\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4fdd\u7559\u4e86\u539f\u59cb\u8f83\u5927\u6a21\u578b\u7684\u5927\u90e8\u5206\u9884\u6d4b\u80fd\u529b\uff0c\u540c\u65f6\u8fd0\u884c\u901f\u5ea6\u66f4\u5feb\u4e14\u8d44\u6e90\u6d88\u8017\u66f4\u5c11\u3002<\/p>\n<p>\u84b8\u998f\u65b9\u5f0f\u4e3b\u8981\u5305\u62ec\u4e24\u79cd\uff1aSDG \u5fae\u8c03\u4e0e\u7ecf\u5178<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\uff0c\u8fd9\u4e24\u79cd\u84b8\u998f\u65b9\u5f0f\u4e92\u8865\u3002\u672c\u6587\u4e3b\u8981\u5173\u6ce8\u7ecf\u5178<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\u65b9\u6cd5\u3002<\/p>\n<p>\u82f1\u4f1f\u8fbe\u91c7\u7528\u5c06<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e0e\u7ecf\u5178<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\u76f8\u7ed3\u5408\u7684\u65b9\u5f0f\u6765\u6784\u9020\u5927\u6a21\u578b\uff0c\u4e0b\u56fe\u5c55\u793a\u4e86\u5355\u4e2a\u6a21\u578b\u7684<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\u8fc7\u7a0b\uff08\u4e0a\uff09\u4ee5\u53ca\u6a21\u578b<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\u7684\u94fe\u6761\uff08\u4e0b\uff09\u3002\u5177\u4f53\u8fc7\u7a0b\u5982\u4e0b\uff1a<\/p>\n<p>1. \u82f1\u4f1f\u8fbe\u4ece 15B \u6a21\u578b\u5f00\u59cb\uff0c\u8bc4\u4f30\u6bcf\u4e2a\u7ec4\u4ef6\uff08\u5c42\u3001<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u3001\u5934\u548c\u5d4c\u5165\u901a\u9053\uff09\u7684\u91cd\u8981\u6027\uff0c\u7136\u540e\u5bf9\u6a21\u578b\u8fdb\u884c\u6392\u5e8f\u548c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\uff0c\u4f7f\u5176\u8fbe\u5230\u76ee\u6807\u5927\u5c0f\uff1a8B \u6a21\u578b\u3002<\/p>\n<p>2. \u63a5\u7740\u4f7f\u7528\u6a21\u578b\u84b8\u998f\u8fdb\u884c\u4e86\u8f7b\u5ea6\u518d\u8bad\u7ec3\uff0c\u539f\u59cb\u6a21\u578b\u4f5c\u4e3a\u8001\u5e08\uff0c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u540e\u7684\u6a21\u578b\u4f5c\u4e3a\u5b66\u751f\u3002<\/p>\n<p>3. \u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u4ee5\u5c0f\u6a21\u578b\uff088B\uff09\u4e3a\u8d77\u70b9\uff0c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\u4e3a\u66f4\u5c0f\u7684 4B \u6a21\u578b\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9d33f4327837db45205f7db36253c921.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9d33f4327837db45205f7db36253c921.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4ece 15B \u6a21\u578b\u8fdb\u884c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e0e\u84b8\u998f\u7684\u8fc7\u7a0b\u3002<\/sup><\/em><\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u70b9\u662f\uff0c\u5728\u5bf9\u6a21\u578b<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e4b\u524d\uff0c\u9700\u8981\u5148\u4e86\u89e3\u6a21\u578b\u7684\u54ea\u90e8\u5206\u662f\u91cd\u8981\u7684\u3002\u82f1\u4f1f\u8fbe\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6fc0\u6d3b\u7684\u7eaf\u91cd\u8981\u6027\u8bc4\u4f30\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u53ef\u4ee5\u540c\u65f6\u8ba1\u7b97\u6240\u6709\u76f8\u5173\u7ef4\u5ea6\uff08\u6df1\u5ea6\u3001<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u3001\u5934\u548c\u5d4c\u5165\u901a\u9053\uff09\u7684\u4fe1\u606f\uff0c\u4f7f\u7528\u4e00\u4e2a\u5305\u542b 1024 \u4e2a\u6837\u672c\u7684\u5c0f\u578b\u6821\u51c6\u6570\u636e\u96c6\uff0c\u5e76\u4e14\u53ea\u9700\u8981\u524d\u5411\u4f20\u64ad\u3002\u8fd9\u79cd\u65b9\u6cd5\u76f8\u6bd4\u4f9d\u8d56\u68af\u5ea6\u4fe1\u606f\u5e76\u9700\u8981\u53cd\u5411\u4f20\u64ad\u7684\u7b56\u7565\u66f4\u52a0\u7b80\u5355\u4e14\u5177\u6709\u6210\u672c\u6548\u76ca\u3002\u00a0<\/p>\n<p>\u5728<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u8fc7\u7a0b\u4e2d\uff0c\u4f60\u53ef\u4ee5\u9488\u5bf9\u7ed9\u5b9a\u8f74\u6216\u8f74\u7ec4\u5408\u5728<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u91cd\u8981\u6027\u4f30\u8ba1\u4e4b\u95f4\u8fdb\u884c\u8fed\u4ee3\u4ea4\u66ff\u3002\u5b9e\u8bc1\u7814\u7a76\u663e\u793a\uff0c\u4f7f\u7528\u5355\u6b21\u91cd\u8981\u6027\u4f30\u8ba1\u5c31\u8db3\u591f\u4e86\uff0c\u8fed\u4ee3\u4f30\u8ba1\u4e0d\u4f1a\u5e26\u6765\u989d\u5916\u7684\u597d\u5904\u3002<\/p>\n<p><strong>\u5229\u7528\u7ecf\u5178<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\u8fdb\u884c\u91cd\u65b0\u8bad\u7ec3<\/strong><\/p>\n<p>\u4e0b\u56fe 2 \u5c55\u793a\u4e86\u84b8\u998f\u8fc7\u7a0b\uff0c\u5176\u4e2d N \u5c42\u5b66\u751f\u6a21\u578b\uff08<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u540e\u7684\u6a21\u578b\uff09\u662f\u4ece M \u5c42\u6559\u5e08\u6a21\u578b\u4e2d\uff08\u539f\u59cb\u672a<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u6a21\u578b\uff09\u84b8\u998f\u800c\u6765\u3002\u5b66\u751f\u6a21\u578b\u901a\u8fc7\u6700\u5c0f\u5316\u5d4c\u5165\u8f93\u51fa\u635f\u5931\u3001logit \u635f\u5931\u4ee5\u53ca<mark data-type=\"concepts\" data-id=\"8ec6a68f-ad96-4b85-ab72-6f8931886922\">\u6620\u5c04<\/mark>\u5230\u5b66\u751f\u5757 S \u548c\u6559\u5e08\u5757 T \u7684 Transformer \u7f16\u7801\u5668\u7279\u5b9a\u635f\u5931\u7ec4\u5408\u6765\u5b66\u4e60\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4ca99c99b4cf791f8d4ffc13ea46d940.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4ca99c99b4cf791f8d4ffc13ea46d940.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 2\uff1a\u84b8\u998f\u8bad\u7ec3\u635f\u5931\u3002<\/sup><\/em><\/p>\n<p><strong><mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\u6700\u4f73\u5b9e\u8df5<\/strong><\/p>\n<p>\u82f1\u4f1f\u8fbe\u57fa\u4e8e\u7d27\u51d1<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4e2d<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c<mark data-type=\"tech_methods\" data-id=\"da9f445a-9f12-49f7-a458-cbacfc40528f\">\u77e5\u8bc6\u84b8\u998f<\/mark>\u7684\u5e7f\u6cdb\u6d88\u878d\u7814\u7a76\uff0c\u5c06\u81ea\u5df1\u7684\u5b66\u4e60\u6210\u679c\u603b\u7ed3\u4e3a\u4ee5\u4e0b\u51e0\u79cd\u7ed3\u6784\u5316\u538b\u7f29\u6700\u4f73\u5b9e\u8df5\u3002<\/p>\n<p>\u4e00\u662f\u8c03\u6574\u5927\u5c0f\u3002<\/p>\n<ul>\n<li>\n<p>\u8981\u8bad\u7ec3\u4e00\u7ec4 LLM\uff0c\u9996\u5148\u8bad\u7ec3\u6700\u5927\u7684\u4e00\u4e2a\uff0c\u7136\u540e\u8fed\u4ee3\u5730<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u84b8\u998f\u4ee5\u83b7\u5f97\u8f83\u5c0f\u7684 LLM\u3002<\/p>\n<\/li>\n<li>\n<p>\u5982\u679c\u4f7f\u7528\u591a\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565\u6765\u8bad\u7ec3\u6700\u5927\u7684\u6a21\u578b\uff0c\u6700\u597d<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u5e76\u5bf9\u8bad\u7ec3\u6700\u540e\u9636\u6bb5\u83b7\u5f97\u7684\u6a21\u578b\u8fdb\u884c\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<\/li>\n<li>\n<p>\u5bf9\u6700\u63a5\u8fd1\u76ee\u6807\u5927\u5c0f\u7684\u53ef\u7528\u6e90\u6a21\u578b\u8fdb\u884c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e8c\u662f<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u3002<\/p>\n<ul>\n<li>\n<p>\u4f18\u5148\u8003\u8651\u5bbd\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u800c\u4e0d\u662f\u6df1\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\uff0c\u8fd9\u5bf9\u4e8e 15B <mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u89c4\u6a21\u4ee5\u4e0b\u7684\u6a21\u578b\u6548\u679c\u5f88\u597d\u3002<\/p>\n<\/li>\n<li>\n<p>\u4f7f\u7528\u5355\u6837\u672c\uff08single-shot\uff09\u91cd\u8981\u6027\u4f30\u8ba1\uff0c\u56e0\u4e3a\u8fed\u4ee3\u91cd\u8981\u6027\u4f30\u8ba1\u6ca1\u6709\u4efb\u4f55\u597d\u5904\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e09\u662f\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<ul>\n<li>\n<p>\u4ec5\u4f7f\u7528\u84b8\u998f\u635f\u5931\u8fdb\u884c\u91cd\u65b0\u8bad\u7ec3\uff0c\u800c\u4e0d\u662f\u5e38\u89c4\u8bad\u7ec3\u3002<\/p>\n<\/li>\n<li>\n<p>\u5f53\u6df1\u5ea6\u660e\u663e\u51cf\u5c11\u65f6\uff0c\u4f7f\u7528 logit\u3001\u4e2d\u95f4\u72b6\u6001\u548c\u5d4c\u5165\u84b8\u998f\u3002<\/p>\n<\/li>\n<li>\n<p>\u5f53\u6df1\u5ea6\u6ca1\u6709\u660e\u663e\u51cf\u5c11\u65f6\uff0c\u4f7f\u7528 logit-only \u84b8\u998f\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>Llama-3.1-Minitron\uff1a\u5c06\u6700\u4f73\u5b9e\u8df5\u4ed8\u8bf8\u5e94\u7528<\/strong><\/p>\n<p>Meta \u6700\u8fd1\u63a8\u51fa\u4e86\u529f\u80fd\u5f3a\u5927\u7684 Llama 3.1 \u5f00\u6e90\u6a21\u578b\u7cfb\u5217\uff0c\u5728\u8bb8\u591a<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u4e2d\u53ef\u4e0e\u95ed\u6e90\u6a21\u578b\u76f8\u5ab2\u7f8e\u3002Llama 3.1 \u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u8303\u56f4\u4ece\u5de8\u5927\u7684 405B \u5230 70B\u30018B\u3002<\/p>\n<p>\u51ed\u501f Nemotron \u84b8\u998f\u7684\u7ecf\u9a8c\uff0c\u82f1\u4f1f\u8fbe\u7740\u624b\u5c06 Llama 3.1 8B \u6a21\u578b\u84b8\u998f\u4e3a\u66f4\u5c0f\u3001\u66f4\u9ad8\u6548\u7684 4B \u6a21\u578b\uff0c\u91c7\u53d6\u4ee5\u4e0b\u63aa\u65bd\uff1a<\/p>\n<ul>\n<li>\n<p>\u6559\u5e08\u5fae\u8c03<\/p>\n<\/li>\n<li>\n<p>Depth-only <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark><\/p>\n<\/li>\n<li>\n<p>Width-only <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark><\/p>\n<\/li>\n<li>\n<p><mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark><mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark><\/p>\n<\/li>\n<li>\n<p>\u6027\u80fd<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark><\/p>\n<\/li>\n<\/ul>\n<p><strong>\u6559\u5e08\u5fae\u8c03<\/strong><\/p>\n<p>\u4e3a\u4e86\u7ea0\u6b63\u6a21\u578b\u8bad\u7ec3\u6240\u57fa\u4e8e\u7684\u539f\u59cb\u6570\u636e\u96c6\u7684\u5206\u5e03\u504f\u5dee\uff0c\u82f1\u4f1f\u8fbe\u9996\u5148\u5728\u4ed6\u4eec\u7684\u6570\u636e\u96c6\u4e0a\uff0894B token\uff09\u5bf9\u672a<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u7684 8B \u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u8c03\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u5982\u679c\u4e0d\u7ea0\u6b63\u5206\u5e03\u504f\u5dee\uff0c\u6559\u5e08\u6a21\u578b\u5728\u84b8\u998f\u65f6\u4f1a\u4e3a\u6570\u636e\u96c6\u63d0\u4f9b\u6b21\u4f18\u6307\u5bfc\u3002<\/p>\n<p><strong>Depth-only <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark><\/strong><\/p>\n<p>\u4e3a\u4e86\u4ece 8B \u964d\u5230 4B\uff0c\u82f1\u4f1f\u8fbe<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e86 16 \u5c42\uff0850%\uff09\u3002\u4ed6\u4eec\u9996\u5148\u901a\u8fc7\u4ece\u6a21\u578b\u4e2d\u5220\u9664\u6bcf\u4e2a\u5c42\u6216\u8fde\u7eed\u5b50\u5c42\u7ec4\u6765\u8bc4\u4f30\u5b83\u4eec\u7684\u91cd\u8981\u6027\uff0c\u5e76\u89c2\u5bdf\u4e0b\u6e38\u4efb\u52a1\u4e2d LM \u635f\u5931\u7684\u589e\u52a0\u6216<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u7684\u964d\u4f4e\u3002<\/p>\n<p>\u4e0b\u56fe 5 \u663e\u793a\u4e86\u5220\u9664 1\u30012\u30018 \u6216 16 \u5c42\u540e<mark data-type=\"concepts\" data-id=\"698eae06-517e-4332-ba61-2de17c83ffb6\">\u9a8c\u8bc1\u96c6<\/mark>\u4e0a\u7684 LM \u635f\u5931\u503c\u3002\u4f8b\u5982\uff0c\u7b2c 16 \u5c42\u7684\u7ea2\u8272\u56fe\u8868\u793a\u5982\u679c\u5220\u9664\u524d 16 \u5c42\uff0c\u5219\u51fa\u73b0 LM \u635f\u5931\u3002\u7b2c 17 \u5c42\u8868\u793a\u5982\u679c\u4fdd\u7559\u7b2c\u4e00\u5c42\u5e76\u5220\u9664\u7b2c 2 \u81f3\u7b2c 17 \u5c42\uff0c\u4e5f\u51fa\u73b0 LM \u635f\u5931\u3002\u82f1\u4f1f\u8fbe\u89c2\u5bdf\u5230\uff1a\u5f00\u59cb\u548c\u7ed3\u675f\u7684\u5c42\u662f\u6700\u91cd\u8981\u7684\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-8a3c1e99e5c6944f1923f2da9cc4dd18.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-8a3c1e99e5c6944f1923f2da9cc4dd18.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 5\uff1adepth-only <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e2d\u5c42\u7684\u91cd\u8981\u6027\u3002<\/sup><\/em><\/p>\n<p>\u7136\u800c\uff0c\u82f1\u4f1f\u8fbe\u89c2\u5bdf\u5230\uff0c\u8fd9\u79cd LM \u635f\u5931\u4e0d\u4e00\u5b9a\u4e0e\u4e0b\u6e38\u6027\u80fd\u76f4\u63a5\u76f8\u5173\u3002<\/p>\n<p>\u4e0b\u56fe 6 \u663e\u793a\u4e86\u6bcf\u4e2a<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u6a21\u578b\u7684 Winogrande <mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\uff0c\u5b83\u8868\u660e\u6700\u597d\u5220\u9664\u7b2c 16 \u5230\u7b2c 31 \u5c42\uff0c\u5176\u4e2d\u7b2c 31 \u5c42\u662f\u5012\u6570\u7b2c\u4e8c\u5c42\uff0c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u6a21\u578b\u7684 5-shot <mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u660e\u663e\u9ad8\u4e8e\u968f\u673a<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark> (0.5)\u3002\u82f1\u4f1f\u8fbe\u91c7\u7eb3\u4e86\u8fd9\u4e00\u89c1\u89e3\uff0c\u5220\u9664\u4e86\u7b2c 16 \u5230\u7b2c 31 \u5c42\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-1ee9cae32b9ac86ae4e8524d503b927b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-1ee9cae32b9ac86ae4e8524d503b927b.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 6\uff1a\u5f53\u5220\u9664 16 \u5c42\u65f6\uff0c\u5728 Winogrande \u4efb\u52a1\u4e0a\u7684<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u3002<\/sup><\/em><\/p>\n<p><strong>Width-only <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark><\/strong><\/p>\n<p>\u82f1\u4f1f\u8fbe\u6cbf\u5bbd\u5ea6\u8f74<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e86\u5d4c\u5165\uff08\u9690\u85cf\uff09\u548c MLP \u4e2d\u95f4\u7ef4\uff0c\u4ee5\u538b\u7f29 Llama 3.1 8B\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u4f7f\u7528\u524d\u9762\u63cf\u8ff0\u7684\u57fa\u4e8e\u6fc0\u6d3b\u7684\u7b56\u7565\u6765\u8ba1\u7b97\u6bcf\u4e2a\u6ce8\u610f\u5934\u3001\u5d4c\u5165\u901a\u9053\u548c MLP \u9690\u85cf\u7ef4\u5ea6\u7684\u91cd\u8981\u6027\u5206\u6570\u3002<\/p>\n<p>\u5728\u91cd\u8981\u6027\u4f30\u8ba1\u4e4b\u540e\uff0c\u82f1\u4f1f\u8fbe\u9009\u62e9<\/p>\n<ul>\n<li>\n<p>\u5c06 MLP \u4e2d\u95f4\u7ef4\u4ece 14336 <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u5230 9216\u3002<\/p>\n<\/li>\n<li>\n<p>\u5c06\u9690\u85cf\u5927\u5c0f\u4ece 4096 <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u5230 3072\u3002<\/p>\n<\/li>\n<li>\n<p>\u91cd\u65b0\u8bad\u7ec3\u6ce8\u610f\u5934\u6570\u91cf\u548c\u5c42\u6570\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u5728\u5355\u6837\u672c<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u4e4b\u540e\uff0c\u5bbd\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u7684 LM \u635f\u5931\u9ad8\u4e8e\u6df1\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u3002\u7136\u800c\uff0c\u7ecf\u8fc7\u77ed\u6682\u7684\u91cd\u65b0\u8bad\u7ec3\u540e\uff0c\u8d8b\u52bf\u53d1\u751f\u4e86\u9006\u8f6c\u3002<\/p>\n<p><strong><mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark><mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark><\/strong><\/p>\n<p>\u82f1\u4f1f\u8fbe\u4f7f\u7528\u4ee5\u4e0b<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5bf9\u6a21\u578b\u8fdb\u884c\u84b8\u998f<\/p>\n<ul>\n<li>\n<p>\u5cf0\u503c<mark data-type=\"concepts\" data-id=\"444ad8d1-7262-46c9-94b7-fe0aa9227006\">\u5b66\u4e60\u7387<\/mark> = 1e-4<\/p>\n<\/li>\n<li>\n<p>\u6700\u5c0f<mark data-type=\"concepts\" data-id=\"444ad8d1-7262-46c9-94b7-fe0aa9227006\">\u5b66\u4e60\u7387<\/mark> = 1e-5<\/p>\n<\/li>\n<li>\n<p>40 \u6b65\u7ebf\u6027\u9884\u70ed<\/p>\n<\/li>\n<li>\n<p>\u4f59\u5f26\u8870\u51cf\u8ba1\u5212<\/p>\n<\/li>\n<li>\n<p>\u5168\u5c40\u6279\u91cf\u5927\u5c0f = 1152<\/p>\n<\/li>\n<\/ul>\n<p>\u4e0b\u8868 1 \u663e\u793a\u4e86 Llama-3.1-Minitron 4B \u6a21\u578b\u53d8\u4f53\uff08\u5bbd\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u6df1\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\uff09\u4e0e\u539f\u59cb Llama 3.1 8B \u6a21\u578b\u3001\u5176\u4ed6\u7c7b\u4f3c\u5927\u5c0f\u7684\u6a21\u578b\u5728\u8de8\u591a\u4e2a\u9886\u57df\u7684<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u4e2d\u7684\u6027\u80fd\u6bd4\u8f83\u3002\u603b\u4f53\u800c\u8a00\uff0c\u82f1\u4f1f\u8fbe\u518d\u6b21\u8bc1\u5b9e\u4e86\u5bbd\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u7b56\u7565\u76f8\u8f83\u4e8e\u9075\u5faa\u6700\u4f73\u5b9e\u8df5\u7684\u6df1\u5ea6<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u7684\u6709\u6548\u6027\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-77075addd53cc13580d05fd865a73719.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-77075addd53cc13580d05fd865a73719.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 1\uff1aMinitron 4B base \u6a21\u578b\u76f8\u8f83\u4e8e\u7c7b\u4f3c\u89c4\u6a21 base \u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u6bd4\u8f83\u3002<\/sup><\/em><\/p>\n<p>\u4e3a\u4e86\u9a8c\u8bc1\u84b8\u998f\u540e\u7684\u6a21\u578b\u662f\u5426\u53ef\u4ee5\u6210\u4e3a\u5f3a\u5927\u7684\u6307\u4ee4\u6a21\u578b\uff0c\u82f1\u4f1f\u8fbe\u4f7f\u7528 NeMo-Aligner \u5bf9 Llama-3.1-Minitron 4B \u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u8c03\u3002<\/p>\n<p>\u4ed6\u4eec\u4f7f\u7528\u4e86 Nemotron-4 340B \u7684\u8bad\u7ec3\u6570\u636e\uff0c\u5728 IFEval\u3001MT-Bench\u3001ChatRAG-Bench \u548c Berkeley Function Calling Leaderboard (BFCL) \u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u4ee5\u6d4b\u8bd5\u6307\u4ee4\u9075\u5faa\u3001\u89d2\u8272\u626e\u6f14\u3001RAG \u548c\u51fd\u6570\u8c03\u7528\u529f\u80fd\u3002\u6700\u540e\u786e\u8ba4 Llama-3.1-Minitron 4B \u6a21\u578b\u53ef\u4ee5\u6210\u4e3a\u53ef\u9760\u7684\u6307\u4ee4\u6a21\u578b\uff0c\u5176\u8868\u73b0\u4f18\u4e8e\u5176\u4ed6\u57fa\u7ebf SLM\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-7d1ed38fd744cff1aff7acdc35a660f6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-7d1ed38fd744cff1aff7acdc35a660f6.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 2\uff1a\u5bf9\u9f50 Minitron 4B base \u6a21\u578b\u4e0e\u7c7b\u4f3c\u89c4\u6a21\u7684\u5bf9\u9f50\u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u6bd4\u8f83\u3002<\/sup><\/em><\/p>\n<p><strong>\u6027\u80fd<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark><\/strong><\/p>\n<p>\u82f1\u4f1f\u8fbe\u5229\u7528 NVIDIA TensorRT-LLM\uff08\u4e00\u79cd\u7528\u4e8e\u4f18\u5316 LLM \u63a8\u7406\u7684\u5f00\u6e90\u5de5\u5177\u5305\uff09\u4f18\u5316\u4e86 Llama 3.1 8B \u548c Llama-3.1-Minitron 4B \u6a21\u578b\u3002<\/p>\n<p>\u4e0b\u4e24\u5f20\u56fe\u663e\u793a\u4e86\u4e0d\u540c\u6a21\u578b\u5728\u4e0d\u540c\u7528\u4f8b\u4e0b\u4ee5 FP8 \u548c FP16 \u7cbe\u5ea6\u6bcf\u79d2\u7684\u541e\u5410\u91cf\u8bf7\u6c42\uff0c\u8868\u793a\u4e3a 8B \u6a21\u578b\u7684 batch size \u4e3a 32 \u7684\u8f93\u5165\u5e8f\u5217\u957f\u5ea6 \/ \u8f93\u51fa\u5e8f\u5217\u957f\u5ea6 (ISL\/OSL) \u7ec4\u5408\u4ee5\u53ca 4B \u6a21\u578b\u7684 batch size \u4e3a 64 \u7684\u8f93\u5165\u5e8f\u5217\u957f\u5ea6 \/ \u8f93\u51fa\u5e8f\u5217\u957f\u5ea6 (ISL\/OSL) \u7ec4\u5408\uff0c\u8fd9\u8981\u5f52\u529f\u4e8e\u5728\u4e00\u5757\u82f1\u4f1f\u8fbe H100 80GB GPU \u4e0a\uff0c\u8f83\u5c0f\u7684<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u5141\u8bb8\u8f83\u5927\u7684 batch size\u3002<\/p>\n<p>Llama-3.1-Minitron-4B-Depth-Base \u53d8\u4f53\u662f\u6700\u5feb\u7684\uff0c\u5e73\u5747\u541e\u5410\u91cf\u7ea6\u4e3a Llama 3.1 8B \u7684 2.7 \u500d\uff0c\u800c Llama-3.1-Minitron-4B-Width-Base \u53d8\u4f53\u7684\u5e73\u5747\u541e\u5410\u91cf\u7ea6\u4e3a Llama 3.1 8B \u7684 1.8 \u500d\u3002\u4e0e BF16 \u76f8\u6bd4\uff0c\u5728 FP8 \u4e2d\u90e8\u7f72\u8fd8\u53ef\u4f7f\u8fd9\u4e09\u79cd\u578b\u53f7\u7684\u6027\u80fd\u63d0\u9ad8\u7ea6 1.3 \u500d\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-be30e9b03951712763d608b5e58fdaca.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-be30e9b03951712763d608b5e58fdaca.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-3e3b569135d3dbf013202df6a7d0cf50.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-3e3b569135d3dbf013202df6a7d0cf50.png\"><\/a><\/section>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 8\uff1a\u7ec4\u5408\uff1aLlama 3.1 8B \u4e3a BS=32\uff0cLlama-3.1-Minitron 4B \u578b\u53f7\u4e3a BS=64\u30021x H100 80GB GPU\u3002<\/sup><\/em><\/p>\n<p><strong>\u7ed3\u8bba<\/strong><\/p>\n<p><mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u548c\u7ecf\u5178\u77e5\u8bc6\u63d0\u70bc\u662f\u4e00\u79cd\u975e\u5e38\u7ecf\u6d4e\u9ad8\u6548\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u9010\u6b65\u83b7\u5f97\u66f4\u5c0f\u5c3a\u5bf8\u7684 LLM\uff0c\u4e0e\u5728\u6240\u6709\u9886\u57df\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u76f8\u6bd4\uff0c\u53ef\u5b9e\u73b0\u66f4\u9ad8\u7684\u51c6\u786e\u6027\u3002\u4e0e\u5408\u6210\u6570\u636e\u5f0f\u5fae\u8c03\u6216\u4ece\u5934\u5f00\u59cb\u9884\u8bad\u7ec3\u76f8\u6bd4\uff0c\u8fd9\u662f\u4e00\u79cd\u66f4\u6709\u6548\u4e14\u6570\u636e\u6548\u7387\u66f4\u9ad8\u7684\u65b9\u6cd5\u3002<\/p>\n<p>Llama-3.1-Minitron 4B \u662f\u82f1\u4f1f\u8fbe\u9996\u6b21\u5c1d\u8bd5\u4f7f\u7528\u6700\u5148\u8fdb\u7684\u5f00\u6e90 Llama 3.1 \u7cfb\u5217\u5b8c\u6210\u7684\u63a2\u7d22\u3002\u8981\u5728 NVIDIA NeMo \u4e2d\u4f7f\u7528 Llama-3.1 \u7684 SDG \u5fae\u8c03\uff0c\u53ef\u53c2\u9605 GitHub \u4e0a\u7684 \/sdg-law-title-generation \u90e8\u5206\u3002<\/p>\n<p>\u6709\u5173\u66f4\u591a\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u8d44\u6e90\uff1a<\/p>\n<ul>\n<li>\n<p>https:\/\/arxiv.org\/abs\/2407.14679<\/p>\n<\/li>\n<li>\n<p>https:\/\/github.com\/NVlabs\/Minitron<\/p>\n<\/li>\n<li>\n<p>https:\/\/huggingface.co\/nvidia\/Llama-3.1-Minitron-4B-Width-Base<\/p>\n<\/li>\n<li>\n<p>https:\/\/huggingface.co\/nvidia\/Llama-3.1-Minitron-4B-Depth-Base<\/p>\n<\/li>\n<\/ul>\n<p><sup>\u53c2\u8003\u94fe\u63a5\uff1a<\/sup><\/p>\n<p><em><sup>https:\/\/developer.nvidia.com\/blog\/how-to-prune-and-distill-llama-3-1-8b-to-an-nvidia-llama-3-1-minitron-4b-model\/<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-16-4\" target=\"_blank\">\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a [&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":[71,68],"class_list":["post-35136","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a - \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=35136\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=35136\",\"url\":\"https:\/\/17aitech.com\/?p=35136\",\"name\":\"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=35136#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=35136#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png\",\"datePublished\":\"2024-12-19T00:59:50+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=35136#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=35136\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=35136#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=35136#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a\"}]},{\"@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":"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a - \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=35136","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35136","url":"https:\/\/17aitech.com\/?p=35136","name":"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35136#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35136#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png","datePublished":"2024-12-19T00:59:50+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35136#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35136"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35136#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e5e3b31a379b9a6ec54f2eda3b0400cb.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35136#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u82f1\u4f1f\u8fbe\u73a9\u8f6c\u526a\u679d\u3001\u84b8\u998f\uff1a\u628aLlama 3.1 8B\u53c2\u6570\u51cf\u534a\uff0c\u6027\u80fd\u540c\u5c3a\u5bf8\u66f4\u5f3a"}]},{"@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\/35136","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=35136"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35136\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35136"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35136"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35136"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}