{"id":781,"date":"2024-06-23T23:57:17","date_gmt":"2024-06-23T15:57:17","guid":{"rendered":"http:\/\/ai.gitpp.com\/?p=781"},"modified":"2024-06-23T23:57:17","modified_gmt":"2024-06-23T15:57:17","slug":"%e7%a0%b4%e8%af%91-llm%ef%bc%9a%e4%bb%8e-transformer-%e5%88%b0-quantization","status":"publish","type":"post","link":"http:\/\/ai.gitpp.com\/index.php\/2024\/06\/23\/%e7%a0%b4%e8%af%91-llm%ef%bc%9a%e4%bb%8e-transformer-%e5%88%b0-quantization\/","title":{"rendered":"\u7834\u8bd1 LLM\uff1a\u4ece Transformer \u5230  Quantization"},"content":{"rendered":"\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u672c\u6587\u662f\u201c\u638c\u63e1 LLM\u201d\u7cfb\u5217\u7684\u7b2c\u4e00\u90e8\u5206\uff0c\u5728\u8fd9\u91cc\u6211\u4eec\u5c06\u8ba8\u8bba\u4ec0\u4e48\u662f LLM\u3001\u5b83\u4eec\u7684\u7528\u4f8b\u4ee5\u53ca\u5b83\u4eec\u6267\u884c\u7684\u4efb\u52a1\u3002\u6211\u4eec\u8fd8\u5c06\u89e3\u91ca\u5e95\u5c42\u7684 Transformer \u67b6\u6784\u53ca\u5176\u80cc\u540e\u7684\u76f4\u89c9\u3002\u7136\u540e\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u4e0d\u540c\u7684\u5fae\u8c03\u6280\u672f\uff0c\u4f8b\u5982\u5b8c\u5168\u5fae\u8c03\u548c\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u3002\u6211\u4eec\u8fd8\u5c06\u5411\u60a8\u4ecb\u7ecd\u91cf\u5316\u7684\u6982\u5ff5\uff0c\u8fd9\u5c06\u5e2e\u52a9\u60a8\u8282\u7701\u4e00\u4e9b\u8ba1\u7b97\u8d44\u6e90<img loading=\"lazy\" decoding=\"async\" height=\"432\" width=\"768\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_gif\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGJlASD5bbibratqEPNCWeXyrW1bu6RccWpCwcOMsD8WyeMAZTckBusww\/640?wx_fmt=gif&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1\" alt=\"\u56fe\u7247\"><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4ec0\u4e48\u662f\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\uff1f\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u63ed\u79d8 Transformer\uff1a\u6df1\u5ea6\u63a2\u7d22\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5fae\u8c03\u6280\u672f\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u8bc4\u4f30\u65b9\u6cd5\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u91cf\u5316\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u7ed3\u8bba\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4ec0\u4e48\u662f\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\uff1f<\/h2>\n\n\n\n<p>\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u662f\u4e00\u79cd\u4eba\u5de5\u667a\u80fd\u6280\u672f\uff0c\u53ef\u4ee5\u751f\u6210\u5404\u79cd\u7c7b\u578b\u7684\u5185\u5bb9\uff0c\u5305\u62ec\u6587\u672c\u3001\u56fe\u50cf\u3001\u97f3\u9891\u548c\u5408\u6210\u6570\u636e\u3002\u65b0\u7528\u6237\u754c\u9762\u7684\u7b80\u6d01\u6027\u63a8\u52a8\u4e86\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u7684\u6d41\u884c\uff0c\u4f7f\u7528\u6237\u80fd\u591f\u5728\u51e0\u79d2\u949f\u5185\u521b\u5efa\u9ad8\u8d28\u91cf\u7684\u6587\u672c\u3001\u56fe\u50cf\u548c\u89c6\u9891\u3002<\/p>\n\n\n\n<p>\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u662f\u673a\u5668\u5b66\u4e60\u7684\u4e00\u4e2a\u5206\u652f\uff0c\u5176\u4e2d\u7684\u6a21\u578b\u5df2\u7ecf\u5b66\u4f1a\u4e86\u4eba\u7c7b\u751f\u6210\u7684\u5927\u91cf\u6570\u636e\u80cc\u540e\u7684\u7edf\u8ba1\u6a21\u5f0f\u3002\u4f8b\u5982\uff0c\u7814\u7a76\u4eba\u5458\u5728\u6570\u5468\u6216\u6570\u6708\u7684\u65f6\u95f4\u91cc\u7528\u6570\u4e07\u4ebf\u4e2a\u5355\u8bcd\u8bad\u7ec3\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u4f7f\u7528\u5927\u91cf\u8ba1\u7b97\u80fd\u529b\u6765\u7406\u89e3\u81ea\u7136\u8bed\u8a00\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">LLM \u7684\u7528\u4f8b<\/h3>\n\n\n\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM)\uff0c\u901a\u5e38\u79f0\u4e3a\u201c\u5927\u578b\u8bed\u8a00\u6a21\u578b\u201d\uff0c\u662f\u4e00\u79cd\u5148\u8fdb\u7684\u4eba\u5de5\u667a\u80fd\u6a21\u578b\uff0c\u65e8\u5728\u7406\u89e3\u548c\u751f\u6210\u7c7b\u4f3c\u4eba\u7c7b\u7684\u6587\u672c\u3002\u5b83\u4eec\u63a5\u53d7\u5927\u91cf\u6587\u672c\u6570\u636e\u7684\u8bad\u7ec3\uff0c\u4ee5\u5b66\u4e60\u4eba\u7c7b\u8bed\u8a00\u7684\u7ed3\u6784\u548c\u6a21\u5f0f\u3002LLM \u9664\u4e86\u804a\u5929\u4efb\u52a1\u5916\uff0c\u8fd8\u6709\u5e7f\u6cdb\u7684\u5e94\u7528\u3002\u867d\u7136\u804a\u5929\u673a\u5668\u4eba\u5df2\u7ecf\u5f15\u8d77\u4e86\u5e7f\u6cdb\u5173\u6ce8\uff0c\u4f46 LLM \u5728\u4e0b\u4e00\u4e2a\u5355\u8bcd\u9884\u6d4b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u8fd9\u6784\u6210\u4e86\u5404\u79cd\u529f\u80fd\u7684\u57fa\u7840\u3002LLM \u7684\u4e00\u4e9b\u5e38\u89c1\u7528\u4f8b\u5305\u62ec\uff1a<\/p>\n\n\n\n<p>\u603b\u7ed3\u6587\u672c\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7ffb\u8bd1\u6587\u672c\uff08\u4f8b\u5982\uff0c\u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u4ee3\u7801\uff09\u3002<\/li>\n\n\n\n<li>\u6839\u636e\u7ed9\u5b9a\u7684\u4e3b\u9898\u751f\u6210\u6587\u7ae0\u3002<\/li>\n\n\n\n<li>\u4ece\u6587\u672c\u4e2d\u63d0\u53d6\u7279\u5b9a\u4fe1\u606f\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u9664\u4e86\u8fd9\u4e9b\u5e94\u7528\u4e4b\u5916\uff0c\u589e\u5f3a LLM \u4e5f\u662f\u4e00\u4e2a\u6d3b\u8dc3\u7684\u7814\u7a76\u9886\u57df\uff0c\u91cd\u70b9\u662f\u5c06 LLM \u4e0e\u5916\u90e8\u6570\u636e\u6e90\u8fde\u63a5\u8d77\u6765\u5e76\u8c03\u7528\u5916\u90e8 API\u3002\u8fd9\u79cd\u96c6\u6210\u5141\u8bb8\u6a21\u578b\u5229\u7528\u5728\u9884\u8bad\u7ec3\u671f\u95f4\u53ef\u80fd\u672a\u5b66\u5230\u7684\u4fe1\u606f\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u589e\u5f3a\u5176\u529f\u80fd\u3002\u6211\u4eec\u5c06\u5728\u5373\u5c06\u53d1\u5e03\u7684\u535a\u5ba2\u4e2d\u8ba8\u8bba\u8fd9\u4e9b\u6280\u672f\u3002<\/p>\n\n\n\n<p>LLM \u4e3a\u81ea\u7136\u8bed\u8a00\u751f\u6210\u63d0\u4f9b\u4e86\u5f3a\u5927\u7684\u5de5\u5177\uff0c\u5e76\u6709\u53ef\u80fd\u901a\u8fc7\u81ea\u52a8\u6267\u884c\u9700\u8981\u8bed\u8a00\u7406\u89e3\u548c\u751f\u6210\u7684\u4efb\u52a1\u5f7b\u5e95\u6539\u53d8\u5404\u4e2a\u9886\u57df\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_jpg\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGeCUIiaXRjfc3e74IibRqSRTSo7icxcBlo9HlgAa2At3kwYOSOhZG8SgtQ\/640?wx_fmt=jpeg&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"\u5185\u8054 \u2013 LearnOpenCV \u2013 LearnOpenCV\"\/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u63ed\u79d8 Transformer\uff1a\u6df1\u5ea6\u63a2\u7d22<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48\u662f Transformer\uff1f<\/h3>\n\n\n\n<p>\u4f60\u53ef\u80fd\u4f1a\u60f3\uff0cLLM \u662f\u5982\u4f55\u5b66\u5f97\u8fd9\u4e48\u597d\u7684\uff1fLLM \u6700\u91cd\u8981\u7684\u56e0\u7d20\u662f\u4ec0\u4e48\uff1f\u7b54\u6848\u662f\u8bba\u6587\u300a\u6ce8\u610f\u529b\u5c31\u662f\u4f60\u6240\u9700\u8981\u7684\u4e00\u5207\u300b\u4e2d\u4ecb\u7ecd\u7684 Transformer \u67b6\u6784\u3002\u5f53\u7136\uff0c\u8fd8\u6709\u5176\u4ed6\u5143\u7d20\u4e5f\u540c\u6837\u91cd\u8981\uff0c\u6bd4\u5982 LLM \u4f7f\u7528\u7684\u6d77\u91cf\u6570\u636e\u548c\u7528\u4e8e\u8bad\u7ec3 LLM \u7684\u8bad\u7ec3\u7b97\u6cd5\uff0c\u4f46 Transformer \u67b6\u6784\u662f\u8ba9\u73b0\u4ee3 LLM \u80fd\u591f\u5982\u6b64\u5f88\u597d\u5730\u5229\u7528\u5176\u4ed6\u7ec4\u4ef6\u7684\u5173\u952e\u7ec4\u4ef6\u3002<\/p>\n\n\n\n<p>\u4e0e\u4e4b\u524d\u7528\u4e8e\u751f\u6210\u5f0f AI \u4efb\u52a1\u7684 RNN \u76f8\u6bd4\uff0c\u4f7f\u7528 Transformer \u53ef\u5927\u5e45\u63d0\u9ad8\u6027\u80fd\u3002Transformer \u6027\u80fd\u5982\u6b64\u63d0\u5347\u7684\u5173\u952e\u5728\u4e8e\u5176\u80fd\u591f\u5b66\u4e60\u53e5\u5b50\u4e2d\u6240\u6709\u5355\u8bcd\u7684\u76f8\u5173\u6027\u548c\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e00\u5b66\u4e60\u8fc7\u7a0b\u901a\u8fc7\u7ed9\u5b9a\u53e5\u5b50\u4e2d\u6bcf\u4e2a\u5355\u8bcd\u4e0e\u5176\u4ed6\u6bcf\u4e2a\u5355\u8bcd\u4e4b\u95f4\u7684\u52a0\u6743\u8fde\u63a5\u53d1\u751f\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u4f1a\u5b66\u4e60\u8fd9\u4e9b\u8fde\u63a5\u7684\u6743\u91cd\uff0c\u5373\u6ce8\u610f\u529b\u6743\u91cd\u3002\u8fd9\u79cd\u80fd\u529b\u901a\u5e38\u88ab\u79f0\u4e3a\u81ea\u6211\u6ce8\u610f\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4ece\u5355\u8bcd\u5230\u5411\u91cf<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGiacrj2a8iaenyIT54EKsMWMvq1PrpHmLzchmztmpJ2el6LIy81PREehA\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"Transformer_architecture \u2013 LearnOpenCV\u00a0\"\/><figcaption class=\"wp-element-caption\">\u56fe 1. \u8bba\u6587\u201cAttention Is All You Need\u201d\u4e2d\u63d0\u51fa\u7684 Transformer \u67b6\u6784\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u5982\u60a8\u6240\u89c1\uff0c\u8be5\u67b6\u6784\u4e3b\u8981\u7531\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7ec4\u6210\u3002\u673a\u5668\u5b66\u4e60\u6a21\u578b\u662f\u5927\u578b\u7edf\u8ba1\u8ba1\u7b97\u5668\uff0c\u5e76\u4e14\u4f7f\u7528\u6570\u5b57\u3002\u56e0\u6b64\uff0c\u4f20\u9012\u6587\u672c\u4e4b\u524d\u7684\u7b2c\u4e00\u6b65\u662f\u6807\u8bb0\u8f93\u5165\u5e8f\u5217\u3002\u6807\u8bb0\u8fc7\u7a0b\u5c06\u6bcf\u4e2a\u5355\u8bcd\u6620\u5c04\u5230\u5355\u8bcd\u8bcd\u5178\u4e2d\u7684\u6570\u5b57\uff0c\u4e00\u4e9b\u6807\u8bb0\u5668\u5c06\u82f1\u8bed\u8bcd\u5178\u4e2d\u7684\u5355\u8bcd\u6620\u5c04\u5230\u6570\u5b57\uff0c\u800c\u4e00\u4e9b\u6807\u8bb0\u5668\u5219\u6620\u5c04\u5355\u8bcd\u7684\u90e8\u5206\u3002\u5173\u952e\u662f\uff0c\u4e00\u65e6\u9009\u62e9\u4e86\u6807\u8bb0\u5668\u8fdb\u884c\u8bad\u7ec3\uff0c\u60a8\u5fc5\u987b\u4f7f\u7528\u76f8\u540c\u7684\u6807\u8bb0\u5668\u8fdb\u884c\u6587\u672c\u751f\u6210\u4efb\u52a1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGPB5SUH7veSvPkY5BmtdtroUGOlQ8ylS4JoZ7uibicMdQjUJyrHia77eGA\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"Token_light \u2013 LearnOpenCV\u00a0\"\/><figcaption class=\"wp-element-caption\">\u56fe 2.\u8f93\u5165\u53e5\u5b50\u7684\u6807\u8bb0\u5316\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u7b2c\u4e8c\u6b65\u662f\u5c06\u8fd9\u4e9b\u6807\u8bb0\u4f20\u9012\u5230\u5d4c\u5165\u5c42\uff0c\u8fd9\u662f\u4e00\u4e2a\u53ef\u8bad\u7ec3\u5c42\uff0c\u8bcd\u6c47\u8868\u4e2d\u7684\u6bcf\u4e2a\u6807\u8bb0\u90fd\u88ab\u5206\u914d\u4e00\u4e2a\u9ad8\u7ef4\u5411\u91cf\uff0c\u8be5\u5411\u91cf\u5b66\u4e60\u5bf9\u8f93\u5165\u5e8f\u5217\u4e2d\u5404\u4e2a\u6807\u8bb0\u7684\u542b\u4e49\u548c\u4e0a\u4e0b\u6587\u8fdb\u884c\u7f16\u7801\u3002\u901a\u8fc7\u5c06\u5355\u8bcd\u7f16\u7801\u4e3a\u9ad8\u7ef4\u5411\u91cf\uff0c\u6a21\u578b\u53ef\u4ee5\u4ece\u6570\u5b66\u4e0a\u5c1d\u8bd5\u7406\u89e3\u8bed\u8a00\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGvPl1B7CfibUstTuBaviagLVTAxCTia2P484jic4F1YlckUMxwamKkdqL7g\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"Embedding_light \u2013 LearnOpenCV\u00a0\"\/><figcaption class=\"wp-element-caption\">\u56fe 3. \u4ee4\u724c\u5d4c\u5165\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u4e0b\u4e00\u6b65\u662f\u5bf9\u8f93\u5165\u5e8f\u5217\u4e2d\u6bcf\u4e2a\u5355\u8bcd\u7684\u4f4d\u7f6e\u8fdb\u884c\u7f16\u7801\u3002\u7ecf\u8fc7\u8fd9\u4e00\u6b65\uff0c\u6765\u81ea\u5d4c\u5165\u5c42\u548c\u4f4d\u7f6e\u7f16\u7801\u5c42\u7684\u5411\u91cf\u88ab\u52a0\u8d77\u6765\u5e76\u4f20\u9012\u5230\u81ea\u6ce8\u610f\u529b\u5c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBG2HjZxaFwOic3dAOhLl5YvecVT06BRRMbY443uKs8FH6icYZdKhm09oWw\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"transformer_lightdrawio \u2013 LearnOpenCV\u00a0\"\/><figcaption class=\"wp-element-caption\">\u56fe 4.\u4f4d\u7f6e\u5d4c\u5165\u3002<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">\u81ea\u6ce8\u610f\u529b\uff1a\u7406\u89e3\u6587\u672c\u4e2d\u7684\u5173\u7cfb<\/h3>\n\n\n\n<p>\u5982\u524d\u6240\u8ff0\uff0c\u81ea\u6ce8\u610f\u529b\u5c42\u5b66\u4e60\u4e00\u4e2a\u5355\u8bcd\u4e0e\u8f93\u5165\u5e8f\u5217\u4e2d\u6240\u6709\u5176\u4ed6\u5355\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u8fd9\u662f\u901a\u8fc7\u4e3a\u8f93\u5165\u5e8f\u5217\u5404\u90e8\u5206\u4e4b\u95f4\u5b58\u5728\u7684\u6240\u6709\u8fde\u63a5\u5206\u914d\u6ce8\u610f\u529b\u6743\u91cd\u6765\u5b9e\u73b0\u7684\u3002<\/p>\n\n\n\n<p>\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u90fd\u7531\u591a\u4e2a\u81ea\u6ce8\u610f\u529b\u5934\u7ec4\u6210\uff0c\u6211\u7684\u610f\u601d\u662f\u8bf4\u6709\u591a\u4e2a\u5c42\uff0c\u6bcf\u4e2a\u5c42\u90fd\u72ec\u7acb\u5b66\u4e60\u8f93\u5165\u5e8f\u5217\u4e2d\u4e0d\u540c\u5355\u8bcd\u4e4b\u95f4\u7684\u6ce8\u610f\u529b\u6743\u91cd\u3002\u8fd9\u91cc\u7684\u76f4\u89c9\u662f\u6bcf\u4e2a\u5934\u53ef\u80fd\u4f1a\u5b66\u4e60\u8bed\u8a00\u7684\u4e0d\u540c\u65b9\u9762\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u5934\u53ef\u80fd\u5173\u6ce8\u53e5\u5b50\u4e2d\u6d89\u53ca\u7684\u60c5\u611f\uff0c\u800c\u53e6\u4e00\u4e2a\u5934\u53ef\u80fd\u5173\u6ce8\u53e5\u5b50\u4e2d\u547d\u540d\u5b9e\u4f53\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u8fd9\u4e9b\u6743\u91cd\u4e2d\u7684\u6bcf\u4e00\u4e2a\u90fd\u662f\u5728\u6a21\u578b\u8bad\u7ec3\u671f\u95f4\u5b66\u4e60\u7684\u3002\u6700\u521d\uff0c\u6743\u91cd\u662f\u968f\u673a\u5206\u914d\u7684\uff0c\u53ea\u8981\u6709\u8db3\u591f\u7684\u6570\u636e\u548c\u65f6\u95f4\uff0c\u8fd9\u4e9b\u6743\u91cd\u5c31\u4f1a\u5b66\u4e60\u8bed\u8a00\u7684\u4e0d\u540c\u65b9\u9762\u3002<\/p>\n\n\n\n<p>\u5728\u7ecf\u8fc7\u591a\u5934\u81ea\u6ce8\u610f\u529b\u5c42\u4e4b\u540e\uff0c\u6ce8\u610f\u529b\u6743\u91cd\u88ab\u5e94\u7528\u5230\u8f93\u5165\u5e8f\u5217\u4e0a\u3002\u968f\u540e\uff0c\u8be5\u5e8f\u5217\u7ecf\u8fc7\u524d\u9988\u7f51\u7edc\u5c42\u7684\u5904\u7406\uff0c\u8be5\u5c42\u751f\u6210\u4e00\u4e2a\u5305\u542b\u8bcd\u6c47\u8868\u4e2d\u6bcf\u4e2a\u5355\u8bcd\u6982\u7387\u7684 logit \u5411\u91cf\u3002\u8be5\u5411\u91cf\u53ef\u4ee5\u4f20\u9012\u5230 softmax \u5c42\u4ee5\u83b7\u53d6\u6bcf\u4e2a\u5355\u8bcd\u7684\u6982\u7387\u5206\u6570\uff0c\u5e76\u53ef\u4ee5\u4f7f\u7528\u5404\u79cd\u65b9\u6cd5\u6765\u9009\u62e9\u6700\u7ec8\u8f93\u51fa\u3002\u8981\u4e86\u89e3\u6ce8\u610f\u529b\u673a\u5236\u7684\u5e95\u5c42\u6570\u5b66\u516c\u5f0f\uff0c\u60a8\u53ef\u4ee5\u67e5\u770b\u300a\u7406\u89e3 Transformer \u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u6ce8\u610f\u529b\u673a\u5236\u300b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7efc\u5408\u8d77\u6765\uff1aTransformer \u6982\u8ff0<\/h3>\n\n\n\n<p>\u73b0\u5728\u6211\u4eec\u4e86\u89e3\u4e86\u5404\u4e2a\u7ec4\u6210\u90e8\u5206\uff0c\u8ba9\u6211\u4eec\u770b\u770b\u7aef\u5230\u7aef\u9884\u6d4b\u8fc7\u7a0b\u3002\u8003\u8651\u4e00\u4e2a\u7ffb\u8bd1\u4efb\u52a1\uff0c\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u5355\u8bcd\u90fd\u88ab\u6807\u8bb0\u5316\u5e76\u901a\u8fc7\u7f16\u7801\u5668\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\u5c42\u8fdb\u884c\u5904\u7406\u3002\u7f16\u7801\u5668\u7684\u8f93\u51fa\u4f1a\u5f71\u54cd\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3002\u6dfb\u52a0\u5e8f\u5217\u8d77\u59cb\u6807\u8bb0\u4ee5\u89e6\u53d1\u4e0b\u4e00\u4e2a\u6807\u8bb0\u7684\u9884\u6d4b\u3002\u8be5\u8fc7\u7a0b\u6301\u7eed\u8fdb\u884c\uff0c\u76f4\u5230\u6ee1\u8db3\u505c\u6b62\u6761\u4ef6\uff0c\u4ece\u800c\u4ea7\u751f\u53ef\u4ee5\u53bb\u6807\u8bb0\u5316\u7684\u6807\u8bb0\u5e8f\u5217\u4ee5\u5f62\u6210\u8f93\u51fa\u5355\u8bcd\u5e8f\u5217\u3002<\/p>\n\n\n\n<p>\u4ece\u9ad8\u5c42\u6b21\u603b\u7ed3\u8be5\u67b6\u6784\uff0c\u7f16\u7801\u5668\u5c06\u8f93\u5165\uff08\u63d0\u793a\uff09\u7f16\u7801\u4e3a\u5bf9\u8f93\u5165\u7684\u7ed3\u6784\u548c\u542b\u4e49\u7684\u6df1\u523b\u7406\u89e3\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u8f93\u5165\u6807\u8bb0\u751f\u6210\u4e00\u4e2a\u5411\u91cf\uff0c\u89e3\u7801\u5668\u63a5\u53d7\u8f93\u5165\u6807\u8bb0\u5e76\u4f7f\u7528\u7f16\u7801\u5668\u63d0\u4f9b\u7684\u7406\u89e3\u751f\u6210\u65b0\u6807\u8bb0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5fae\u8c03\u6280\u672f<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u6307\u4ee4\u5fae\u8c03<\/h3>\n\n\n\n<p>\u4e0e\u901a\u8fc7\u81ea\u76d1\u7763\u5b66\u4e60\u5728\u5927\u91cf\u975e\u7ed3\u6784\u5316\u6587\u672c\u6570\u636e\u4e0a\u8bad\u7ec3 LLM \u7684\u9884\u8bad\u7ec3\u4e0d\u540c\uff0c\u5fae\u8c03\u662f\u4e00\u4e2a\u76d1\u7763\u5b66\u4e60\u8fc7\u7a0b\uff0c\u4f7f\u7528\u5e26\u6709\u6807\u8bb0\u793a\u4f8b\u7684\u6570\u636e\u96c6\u66f4\u65b0\u6743\u91cd\u3002\u8fd9\u6709\u52a9\u4e8e\u60a8\u63d0\u9ad8\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002\u6807\u8bb0\u793a\u4f8b\u53ef\u4ee5\u662f\u60c5\u7eea\u5206\u6790\u4efb\u52a1\u3001\u7ffb\u8bd1\u4efb\u52a1\u3001\u6458\u8981\u4efb\u52a1\u6216\u5176\u4ed6\u4efb\u52a1\u7684\u793a\u4f8b\uff0c\u4e5f\u53ef\u4ee5\u662f\u4e0a\u8ff0\u4efb\u52a1\u7684\u7ec4\u5408\u3002\u6807\u8bb0\u793a\u4f8b\u5305\u542b\u4e24\u90e8\u5206\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u63d0\u793a\uff08\u5bf9\u4e8e\u60c5\u7eea\u5206\u6790\u4efb\u52a1\uff0c\u53ef\u80fd\u662f\uff1a\u6211\u4eca\u5929\u611f\u89c9\u5f88\u597d\u3002\uff09<\/li>\n\n\n\n<li>\u5b8c\u6210\uff08\u5bf9\u4e8e\u4e0a\u8ff0\u63d0\u793a\uff0c\u5176\u7b54\u6848\u662f\uff1a\u79ef\u6781\u7684\u3002\uff09<\/li>\n<\/ul>\n\n\n\n<p>\u8bb8\u591a\u516c\u5f00\u53ef\u7528\u7684\u6570\u636e\u96c6\u53ef\u7528\u4e8e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u4f46\u5176\u4e2d\u5927\u591a\u6570\u6570\u636e\u96c6\u90fd\u4e0d\u662f\u6307\u4ee4\u683c\u5f0f\u3002\u5e78\u8fd0\u7684\u662f\uff0c\u4e00\u4e9b\u5373\u65f6\u6a21\u677f\u5e93\u4e3a\u5404\u79cd\u6570\u636e\u96c6\u7684\u5404\u79cd\u4efb\u52a1\u63d0\u4f9b\u4e86\u6a21\u677f\u3002\u8fd9\u4e9b\u5e93\u53ef\u4ee5\u5904\u7406\u6d41\u884c\u7684\u6570\u636e\u96c6\uff0c\u4ee5\u521b\u5efa\u53ef\u7528\u4e8e\u6307\u4ee4\u5fae\u8c03\u7684\u6570\u636e\u3002<\/p>\n\n\n\n<p>\u6240\u6709\u6a21\u578b\u6743\u91cd\u90fd\u66f4\u65b0\u7684\u6307\u4ee4\u5fae\u8c03\u79f0\u4e3a\u5b8c\u5168\u5fae\u8c03\u3002\u50cf\u8fd9\u6837\u7684\u5fae\u8c03\u9700\u8981\u60a8\u5b58\u50a8\u6240\u6709\u6a21\u578b\u6743\u91cd\uff0c\u56e0\u6b64\u8fd9\u9700\u8981\u4e0e\u9884\u8bad\u7ec3\u671f\u95f4\u6240\u9700\u7684\u76f8\u540c\u6570\u91cf\u7684\u8ba1\u7b97\u80fd\u529b\u3002\u60a8\u53ef\u4ee5\u5728 Fine&nbsp;Tuning T5\uff1aText2Text Transfer Transformer \u4e2d\u8be6\u7ec6\u63a2\u7d22\u6b64\u65b9\u6cd5\uff0c\u5e76\u901a\u8fc7\u6784\u5efa\u4e00\u4e9b\u6709\u8da3\u7684\u5e94\u7528\u7a0b\u5e8f\u6765\u4eb2\u81ea\u52a8\u624b\uff0c\u4ee5\u4f7f\u7528 T5 \u6784\u5efa Stack Overflow \u6807\u7b7e\u751f\u6210\u5668\u548c\u6587\u672c\u6458\u8981\uff1a\u5fae\u8c03\u548c\u6784\u5efa Gradio \u5e94\u7528\u7a0b\u5e8f<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u591a\u4efb\u52a1\u5fae\u8c03<\/h3>\n\n\n\n<p>\u5728\u591a\u4efb\u52a1\u5fae\u8c03\u4e2d\uff0c\u6211\u4eec\u672c\u8d28\u4e0a\u662f\u540c\u65f6\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u5fae\u8c03\u6a21\u578b\u3002\u8fd9\u9700\u8981\u5927\u91cf\u7684\u8bad\u7ec3\u6837\u672c\u3002\u9002\u5408\u591a\u4efb\u52a1\u5fae\u8c03\u7684\u4efb\u52a1\u5305\u62ec\u60c5\u611f\u5206\u6790\u3001\u6587\u672c\u6458\u8981\u3001\u95ee\u7b54\u3001\u673a\u5668\u7ffb\u8bd1\u7b49\u3002<\/p>\n\n\n\n<p>\u591a\u4efb\u52a1\u5fae\u8c03\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u4e00\u7cfb\u5217\u6311\u6218\u4e2d\u5b66\u4e60\u548c\u6982\u62ec\uff0c\u6700\u7ec8\u63d0\u9ad8\u5176\u6574\u4f53\u6027\u80fd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u53c2\u6570\u6709\u6548\u5fae\u8c03\uff08PEFT\uff09<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u4ec0\u4e48\u662f PEFT\uff1f<\/h4>\n\n\n\n<p>LLM \u53ef\u80fd\u975e\u5e38\u5927\uff0c\u6700\u5927\u7684 LLM \u9700\u8981\u6570\u767e GB \u7684\u5b58\u50a8\u7a7a\u95f4\u3002\u4e3a\u4e86\u80fd\u591f\u5bf9 LLM \u8fdb\u884c\u5168\u9762\u5fae\u8c03\uff0c\u60a8\u4e0d\u4ec5\u9700\u8981\u5b58\u50a8\u6240\u6709\u6a21\u578b\u6743\u91cd\uff0c\u8fd8\u9700\u8981\u5b58\u50a8\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u3001\u524d\u5411\u6fc0\u6d3b\u548c\u4e34\u65f6\u72b6\u6001\u3002\u8fd9\u53ef\u80fd\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u80fd\u529b\uff0c\u800c\u83b7\u5f97\u8fd9\u4e9b\u80fd\u529b\u53ef\u80fd\u975e\u5e38\u56f0\u96be\u3002<\/p>\n\n\n\n<p>\u53c2\u6570\u9ad8\u6548\u65b9\u6cd5\u4e0d\u4f1a\u50cf\u5168\u5fae\u8c03\u90a3\u6837\u5728\u8bad\u7ec3\u671f\u95f4\u66f4\u65b0\u6bcf\u4e2a\u6743\u91cd\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\u4e00\u5c0f\u90e8\u5206\u3002\u8fd9\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5fae\u8c03\u7279\u5b9a\u7684\u5c42\u6216\u7ec4\u4ef6\uff1a\u00a0\u53ea\u8c03\u6574\u76f8\u5173\u90e8\u5206\uff0c\u800c\u4e0d\u662f\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u6a21\u578b\u3002<\/li>\n\n\n\n<li>\u51bb\u7ed3 LLM \u6743\u91cd\u5e76\u6dfb\u52a0\u65b0\u7ec4\u4ef6\uff1a\u00a0\u539f\u59cb\u6a21\u578b\u53d8\u4e3a\u53ea\u8bfb\uff0c\u540c\u65f6\u5f15\u5165\u5e76\u4e13\u95e8\u9488\u5bf9\u4efb\u52a1\u8fdb\u884c\u8bad\u7ec3\u7684\u9644\u52a0\u5c42\u6216\u53c2\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u603b\u4f53\u800c\u8a00\uff0c\u4ec5\u66f4\u65b0\u7ea6 15-20\uff05 \u7684\u539f\u59cb\u53c2\u6570\uff0c\u5373\u4f7f\u5728\u8ba1\u7b97\u8d44\u6e90\u8f83\u5c11\u7684\u5355\u4e2a GPU \u4e0a\u4e5f\u53ef\u4ee5\u8fdb\u884c\u53c2\u6570\u9ad8\u6548\u7684\u5fae\u8c03\u3002<\/p>\n\n\n\n<p>\u7531\u4e8e\u5927\u591a\u6570 LLM \u6743\u91cd\u4ec5\u8fdb\u884c\u8f7b\u5fae\u4fee\u6539\u6216\u4fdd\u6301\u4e0d\u53d8\uff0c\u56e0\u6b64\u53d1\u751f\u707e\u96be\u6027\u9057\u5fd8\u7684\u98ce\u9669\u663e\u8457\u964d\u4f4e\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">PEFT \u8bfe\u7a0b<\/h4>\n\n\n\n<p>PEFT \u65b9\u6cd5\u4e3b\u8981\u6709\u4ee5\u4e0b 3 \u4e2a\u7c7b\u522b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u9009\u62e9\u6027\u65b9\u6cd5\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u4e13\u95e8\u7528\u4e8e\u5fae\u8c03\u521d\u59cb LLM \u53c2\u6570\u7684\u4e00\u90e8\u5206\u3002\u60a8\u53ef\u4ee5\u901a\u8fc7\u591a\u79cd\u9009\u9879\u6765\u51b3\u5b9a\u8981\u4fee\u6539\u54ea\u4e9b\u53c2\u6570\u3002\u60a8\u53ef\u4ee5\u9009\u62e9\u8bad\u7ec3\u6a21\u578b\u7684\u7279\u5b9a\u7ec4\u4ef6\u3001\u7279\u5b9a\u5c42\uff0c\u751a\u81f3\u7279\u5b9a\u7c7b\u578b\u7684\u53c2\u6570\u3002<\/li>\n\n\n\n<li>\u91cd\u65b0\u53c2\u6570\u5316\u65b9\u6cd5\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u521b\u5efa\u539f\u59cb\u7f51\u7edc\u6743\u91cd\u7684\u65b0\u4f4e\u79e9\u53d8\u6362\u6765\u51cf\u5c11\u8981\u8bad\u7ec3\u7684\u53c2\u6570\u6570\u91cf\u3002\u8fd9\u79cd\u7c7b\u578b\u7684\u4e00\u79cd\u5e38\u7528\u6280\u672f\u662f LoRA\uff0c\u6211\u4eec\u5c06\u5728 Mastering LLM \u7cfb\u5217\u7684\u4e0b\u4e00\u7bc7\u4e2d\u63a2\u8ba8\u5b83\u3002<\/li>\n\n\n\n<li>\u9644\u52a0\u65b9\u6cd5\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u4fdd\u6301\u539f\u59cb LLM \u6743\u91cd\u4e0d\u53d8\u5e76\u5408\u5e76\u65b0\u7684\u53ef\u8bad\u7ec3\u7ec4\u4ef6\u6765\u6267\u884c\u5fae\u8c03\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6709\u4e24\u79cd\u4e3b\u8981\u65b9\u6cd5\u3002\u9002\u914d\u5668\u65b9\u6cd5\u5c06\u65b0\u7684\u53ef\u8bad\u7ec3\u5c42\u5f15\u5165\u6a21\u578b\u67b6\u6784\u4e2d\uff0c\u901a\u5e38\u5728\u7f16\u7801\u5668\u6216\u89e3\u7801\u5668\u7ec4\u4ef6\u4e2d\uff0c\u7d27\u968f\u6ce8\u610f\u5c42\u6216\u524d\u9988\u5c42\u4e4b\u540e\u3002\u76f8\u53cd\uff0c\u8f6f\u63d0\u793a\u65b9\u6cd5\u4fdd\u6301\u56fa\u5b9a\u548c\u51bb\u7ed3\u7684\u6a21\u578b\u67b6\u6784\uff0c\u540c\u65f6\u4e13\u6ce8\u4e8e\u64cd\u7eb5\u8f93\u5165\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002\u8fd9\u53ef\u4ee5\u901a\u8fc7\u5411\u63d0\u793a\u5d4c\u5165\u6dfb\u52a0\u53ef\u8bad\u7ec3\u53c2\u6570\u6216\u4fdd\u6301\u8f93\u5165\u4e0d\u53d8\u5e76\u91cd\u65b0\u8bad\u7ec3\u5d4c\u5165\u6743\u91cd\u6765\u5b9e\u73b0\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u8bc4\u4f30\u65b9\u6cd5<\/h2>\n\n\n\n<p>\u5728\u4f20\u7edf\u7684\u673a\u5668\u5b66\u4e60\u65b9\u6cd5\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u8bc4\u4f30\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u6216\u9a8c\u8bc1\u96c6\u4e0a\u7684\u8868\u73b0\uff0c\u56e0\u4e3a\u8f93\u51fa\u662f\u786e\u5b9a\u6027\u7684\u3002\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u51c6\u786e\u7387\u3001\u7cbe\u786e\u5ea6\u3001\u53ec\u56de\u7387\u3001f1 \u5206\u6570\u7b49\u6307\u6807\u6765\u786e\u5b9a\u6a21\u578b\u7684\u8868\u73b0\u5982\u4f55\u3002\u7136\u800c\uff0c\u5bf9\u4e8e LLM \u6765\u8bf4\uff0c\u8fd9\u53ef\u80fd\u5f88\u56f0\u96be\uff0c\u56e0\u4e3a\u8f93\u51fa\u662f\u4e0d\u786e\u5b9a\u7684\uff0c\u800c\u57fa\u4e8e\u8bed\u8a00\u7684\u8bc4\u4f30\u53ef\u80fd\u975e\u5e38\u5177\u6709\u6311\u6218\u6027\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\uff0c\u8003\u8651\u4ee5\u4e0b\u53e5\u5b50\uff1a\u201c\u4eca\u5929\u5929\u6c14\u5f88\u597d\u201d\u548c\u201c\u4eca\u5929\u5929\u6c14\u5f88\u597d\u201d\u3002\u4e0a\u8ff0\u4e24\u4e2a\u53e5\u5b50\u5177\u6709\u76f8\u4f3c\u7684\u542b\u4e49\uff0c\u4f46\u6211\u4eec\u5982\u4f55\u8861\u91cf\u8fd9\u79cd\u76f8\u4f3c\u6027\uff1f\u4f8b\u5982\uff0c\u8003\u8651\u4ee5\u4e0b\u53e5\u5b50\uff1a\u201c\u4eca\u5929\u5929\u6c14\u5f88\u7cdf\u7cd5\u201d\uff0c\u73b0\u5728\u8fd9\u4e2a\u53e5\u5b50\u4e0e\u5176\u4ed6\u4e24\u4e2a\u53e5\u5b50\u7684\u542b\u4e49\u76f8\u53cd\u3002\u4f46\u8fd9\u4e09\u4e2a\u53e5\u5b50\u53ea\u6709\u4e00\u4e2a\u8bcd\u7684\u5dee\u5f02\u3002\u6211\u4eec\u4eba\u7c7b\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u533a\u5206\u8fd9\u4e9b\u53e5\u5b50\uff0c\u4f46\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u7ed3\u6784\u5316\u3001\u5b9a\u4e49\u660e\u786e\u7684\u65b9\u6cd5\u6765\u8861\u91cf\u4e0d\u540c\u53e5\u5b50\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u4ee5\u8bad\u7ec3 LLM\u3002ROUGE \u548c BLEU \u662f\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u7684\u4e24\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u8bc4\u4f30\u6307\u6807\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">ROUGE\uff08\u4ee5\u56de\u5fc6\u4e3a\u5bfc\u5411\u7684\u8981\u70b9\u8bc4\u4f30\u66ff\u8865\uff09\u8bc4\u5206<\/h3>\n\n\n\n<p>\u5b83\u901a\u5e38\u7528\u4e8e\u6587\u672c\u6458\u8981\u4efb\u52a1\u3002\u5b83\u7528\u4e8e\u5ba2\u89c2\u8bc4\u4f30\u8f83\u957f\u6587\u672c\u4e2d\u673a\u5668\u751f\u6210\u7684\u6458\u8981\u4e0e\u4eba\u7c7b\u63d0\u4f9b\u7684\u53c2\u8003\u6458\u8981\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u3002\u8ba9\u6211\u4eec\u770b\u770b\u4e00\u4e9b\u7c7b\u578b\u7684 ROUGE \u5206\u6570\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>ROUGE-1 \u5206\u6570\uff1a\u5b83\u6d4b\u91cf\u673a\u5668\u751f\u6210\u7684\u6587\u672c\u548c\u53c2\u8003\u6587\u672c\u4e2d\u5355\u4e2a\u5355\u8bcd\u7684\u91cd\u53e0\u5ea6\u3002\u5b83\u6839\u636e\u91cd\u53e0\u5355\u8bcd\u7684\u6570\u91cf\u8ba1\u7b97\u7cbe\u5ea6\u3001\u53ec\u56de\u7387\u548c F1 \u5206\u6570\u7b49\u91cf\u3002\u867d\u7136\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u8861\u91cf\u4e24\u4e2a\u6587\u672c\u7684\u76f8\u4f3c\u5ea6\uff0c\u4f46\u5b83\u53ef\u80fd\u6709\u70b9\u8bef\u5bfc\u3002\u4f8b\u5982\uff0c\u8003\u8651\u4ee5\u4e0b\u53c2\u8003\u53e5\u5b50\u201c\u8fd9\u91cc\u5f88\u51b7\u201d\uff0c\u7136\u540e\u8003\u8651\u4ee5\u4e0b\u4e24\u4e2a\u673a\u5668\u751f\u6210\u7684\u53e5\u5b50\u201c\u8fd9\u91cc\u5f88\u51b7\u201d\u548c\u201c\u8fd9\u91cc\u4e0d\u51b7\u201d\uff0c\u8fd9\u4e24\u4e2a\u53e5\u5b50\u7684\u6d4b\u91cf\u503c\u5c06\u76f8\u540c\uff0c\u4f46\u662f\uff0c\u5176\u4e2d\u4e00\u4e2a\u4e0e\u53c2\u8003\u6587\u672c\u76f8\u6bd4\u542b\u4e49\u76f8\u53cd\u3002<\/li>\n\n\n\n<li>ROUGE-N \u5206\u6570\uff1a\u5b83\u6d4b\u91cf\u673a\u5668\u751f\u6210\u7684\u6587\u672c\u548c\u53c2\u8003\u6587\u672c\u4e2d n-gram\uff08\u7ed9\u5b9a\u53e5\u5b50\u4e2d\u7684 n \u4e2a\u8fde\u7eed\u5355\u8bcd\uff09\u7684\u91cd\u53e0\u5ea6\u3002\u5b83\u6839\u636e\u91cd\u53e0 n-gram \u7684\u6570\u91cf\u8ba1\u7b97\u51c6\u786e\u7387\u3001\u53ec\u56de\u7387\u548c F1 \u5206\u6570\u7b49\u91cf\u3002\u7531\u4e8e\u6b64\u6307\u6807\u91c7\u7528\u4e8c\u5143\u8bed\u6cd5\u548c n-gram \u8fdb\u884c\u8ba1\u7b97\uff0c\u56e0\u6b64\u76f4\u89c2\u5730\u53ef\u4ee5\u770b\u51fa\u5b83\u4e5f\u8003\u8651\u4e86\u5355\u8bcd\u7684\u4f4d\u7f6e\u3002<\/li>\n\n\n\n<li>ROUGE-L \u5206\u6570\uff1a\u8fd9\u662f\u53e6\u4e00\u79cd\u5e38\u89c1\u7684 ROUGE \u5206\u6570\uff0c\u6211\u4eec\u4f7f\u7528\u6700\u957f\u516c\u5171\u5b50\u5e8f\u5217\u7684\u957f\u5ea6\uff08\u65e0\u8bba\u987a\u5e8f\u5982\u4f55\uff09\u6765\u8ba1\u7b97\u7cbe\u5ea6\u3001\u53ec\u56de\u7387\u548c F1 \u5206\u6570\u7b49\u6570\u91cf\uff0c\u8fd9\u4e9b\u7edf\u79f0\u4e3a ROUGE \u5206\u6570\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">BLEU\uff08\u53cc\u8bed\u8bc4\u4f30\u66ff\u8865\uff09\u5206\u6570<\/h3>\n\n\n\n<p>\u5b83\u662f\u81ea\u52a8\u8bc4\u4f30\u673a\u5668\u7ffb\u8bd1\u6587\u672c\u7684\u6307\u6807\u3002BLEU \u5206\u6570\u662f\u4e00\u4e2a\u4ecb\u4e8e 0 \u548c 1 \u4e4b\u95f4\u7684\u6570\u5b57\uff0c\u7528\u4e8e\u8861\u91cf\u673a\u5668\u7ffb\u8bd1\u6587\u672c\u4e0e\u4e00\u7ec4\u53c2\u8003\u7ffb\u8bd1\u7684\u76f8\u4f3c\u5ea6\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u53ef\u4ee5\u8bf4\u5b83\u662f\u4e00\u7cfb\u5217 n-gram \u5927\u5c0f\u7684\u7cbe\u5ea6\u503c\u7684\u51e0\u4f55\u5e73\u5747\u503c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e00\u4e9b\u5b9a\u6027\u89c1\u89e3<\/h3>\n\n\n\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u5b9a\u91cf\u5730\u770b\u770b\u8fd9\u4e9b\u6307\u6807\u5bf9\u4e8e\u4e0d\u540c\u7684\u6587\u672c\u5bf9\u662f\u5982\u4f55\u53d8\u5316\u7684\u3002\u8003\u8651\u4ee5\u4e0b\u53c2\u8003\u6587\u672c\u3001\u673a\u5668\u751f\u6210\u7684\u6587\u672c\u548c\u76f8\u5e94\u7684\u6307\u6807\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><td>\u53c2\u8003\u6587\u672c<\/td><td>\u4eba\u5de5\u667a\u80fd\u6b63\u5728\u6539\u53d8\u5404\u4e2a\u884c\u4e1a\uff0c\u5305\u62ec\u533b\u7597\u4fdd\u5065\u548c\u91d1\u878d\u3002<\/td><\/tr><tr><td>\u9884\u6d4b 1<\/td><td>\u4eba\u5de5\u667a\u80fd\u6b63\u5728\u5f7b\u5e95\u6539\u53d8\u533b\u7597\u4fdd\u5065\u548c\u91d1\u878d\u7b49\u4e0d\u540c\u9886\u57df\u3002<\/td><\/tr><tr><td>\u9884\u6d4b 2<\/td><td>\u4eba\u5de5\u667a\u80fd\u6b63\u5728\u6539\u53d8\u533b\u7597\u4fdd\u5065\u548c\u91d1\u878d\u7b49\u884c\u4e1a\u3002<\/td><\/tr><tr><td>\u9884\u6d4b 3<\/td><td>\u4eba\u5de5\u667a\u80fd\u6b63\u5728\u6539\u53d8\u5404\u4e2a\u884c\u4e1a\uff0c\u5305\u62ec\u533b\u7597\u4fdd\u5065\u548c\u91d1\u878d\u3002<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u9884\u8a00<\/th><th>BLEU \u5206\u6570<\/th><th>ROUGE-1 \u5206\u6570<\/th><th>ROUGE-2 \u5206\u6570<\/th><th>ROUGE-L \u5206\u6570<\/th><\/tr><\/thead><tbody><tr><td>\u9884\u6d4b 1<\/td><td>0.23708<\/td><td>0.42105<\/td><td>0.23529<\/td><td>0.42105<\/td><\/tr><tr><td>\u9884\u6d4b 2<\/td><td>0.44127<\/td><td>0.8<\/td><td>0.5555<\/td><td>0.8<\/td><\/tr><tr><td>\u9884\u6d4b 3<\/td><td>1.0<\/td><td>1.0<\/td><td>1.0<\/td><td>1.0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u60a8\u53ef\u4ee5\u770b\u5230\uff0c\u968f\u7740\u6211\u4eec\u8d8a\u6765\u8d8a\u63a5\u8fd1\u53c2\u8003\u6587\u672c\uff0cBLEU \u5206\u6570\u548c ROUGE \u5206\u6570\u5982\u4f55\u63d0\u9ad8\u3002\u60a8\u53ef\u4ee5\u4e0b\u8f7d\u4e0b\u9762\u7684\u4ee3\u7801\uff0c\u8bd5\u7528\u53c2\u8003\u6587\u672c\u548c\u9884\u6d4b\u6587\u672c\uff0c\u7136\u540e\u4eb2\u81ea\u770b\u770b\u8fd9\u4e9b\u6307\u6807\u5982\u4f55\u53d8\u5316\u3002<\/p>\n\n\n\n<p>\u4e0b\u8f7d\u4ee3\u7801&nbsp;\u4e3a\u4e86\u8f7b\u677e\u5b66\u4e60\u672c\u6559\u7a0b\uff0c\u8bf7\u70b9\u51fb\u4e0b\u9762\u7684\u6309\u94ae\u4e0b\u8f7d\u4ee3\u7801\u3002\u514d\u8d39\uff01<br><\/p>\n\n\n\n<p>\u4e0b\u8f7d\u4ee3\u7801<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u91cf\u5316<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48\u8981\u91cf\u5316\uff1f<\/h3>\n\n\n\n<p>LLM \u5177\u6709\u6570\u767e\u4e07\u6216\u6570\u5341\u4ebf\u4e2a\u53c2\u6570\uff0c\u6bcf\u4e2a\u53c2\u6570\u90fd\u7528\u4e00\u5b9a\u6570\u91cf\u7684\u4f4d\u8868\u793a\uff0c\u901a\u5e38\u662f 32 \u4f4d\u6216 16 \u4f4d\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u8003\u8651\u4e00\u4e2a\u5177\u6709 10 \u4ebf\u4e2a\u53c2\u6570\u7684\u6a21\u578b\uff0c\u5176\u6240\u6709\u53c2\u6570\u90fd\u7528 32 \u4f4d\u6d6e\u70b9\u6570\u5b58\u50a8\u3002\u6b64\u6a21\u578b\u4e2d\u7684\u6bcf\u4e2a\u53c2\u6570\u5c06\u5360\u7528 4 \u5b57\u8282\u5185\u5b58\u3002\u56e0\u6b64\uff0c\u603b\u5171 10 \u4ebf\u4e2a\u53c2\u6570\u5c06\u5360\u7528 4 \u2217 109 \u5b57\u8282\u5185\u5b58\uff0c\u76f8\u5f53\u4e8e 4GB \u5185\u5b58\u3002<\/p>\n\n\n\n<p>\u8fd9\u610f\u5473\u7740\u8981\u5b58\u50a8\u5177\u6709 10 \u4ebf\u4e2a\u53c2\u6570\u7684 32 \u4f4d\u6a21\u578b\uff0c\u60a8\u9700\u8981 4GB \u5185\u5b58\u3002\u5982\u679c\u60a8\u60f3\u8bad\u7ec3\u8fd9\u6837\u7684\u6a21\u578b\uff0c\u90a3\u4e48\u60a8\u5c06\u9700\u8981\u989d\u5916\u7684\u5185\u5b58\u6765\u5b58\u50a8\u51fd\u6570\u6240\u9700\u7684\u68af\u5ea6\u3001\u6fc0\u6d3b\u3001\u4f18\u5316\u5668\u72b6\u6001\u548c\u5176\u4ed6\u4e34\u65f6\u53d8\u91cf\u3002\u8fd9\u5f88\u5bb9\u6613\u5bfc\u81f4\u6bcf\u4e2a\u6a21\u578b\u53c2\u6570\u9700\u8981\u5927\u7ea6 20 \u5b57\u8282\u7684\u989d\u5916\u5185\u5b58\u3002\u5728\u8bad\u7ec3\u671f\u95f4\u8003\u8651\u5230\u6240\u6709\u8fd9\u4e9b\u989d\u5916\u7684\u5185\u5b58\u9700\u6c42\u540e\uff0c\u60a8\u5c06\u9700\u8981\u5927\u7ea6 6 \u500d\u7684 GPU RAM \u6765\u8bad\u7ec3\u8fd9\u6837\u7684\u6a21\u578b\uff08\u7ea6 24GB\uff09\uff0c\u9664\u6b64\u4e4b\u5916\uff0c\u60a8\u8fd8\u9700\u8981\u4e00\u4e9b\u5185\u5b58\u6765\u5b58\u50a8\u6570\u636e\u3002\u8fd9\u5bf9\u4e8e\u8fb9\u7f18\u8bbe\u5907\u6765\u8bf4\u592a\u591a\u4e86\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u91cf\u5316\u6743\u8861<\/h3>\n\n\n\n<p>\u90a3\u4e48\uff0c\u60a8\u6709\u54ea\u4e9b\u9009\u62e9\u53ef\u4ee5\u51cf\u5c11\u8bad\u7ec3\u6240\u9700\u7684\u5185\u5b58\u5462\uff1f\u60a8\u53ef\u4ee5\u4f7f\u7528\u4e00\u79cd\u51cf\u5c11\u5185\u5b58\u7684\u6280\u672f\uff0c\u79f0\u4e3a\u91cf\u5316\u3002\u901a\u8fc7\u7528 16 \u4f4d\u751a\u81f3 8 \u4f4d\u6574\u6570\u66ff\u6362 32 \u4f4d\u6d6e\u70b9\u6570\uff0c\u91cf\u5316\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\u5927\u578b\u8bed\u8a00\u6a21\u578b\u548c\u795e\u7ecf\u7f51\u7edc\u7684\u5185\u5b58\u9700\u6c42\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u786e\u4fdd\u51c6\u786e\u6027\u548c\u7075\u6d3b\u6027\uff0c\u6a21\u578b\u6743\u91cd\u3001\u6fc0\u6d3b\u548c\u53c2\u6570\u901a\u5e38\u9ed8\u8ba4\u4ee5 32 \u4f4d\u6d6e\u70b9\u683c\u5f0f\u5b58\u50a8\u3002\u91cf\u5316\u4f7f\u7528\u7f29\u653e\u56e0\u5b50\u5c06\u539f\u59cb 32 \u4f4d\u6d6e\u70b9\u6570\u6295\u5f71\u5230\u8f83\u4f4e\u7cbe\u5ea6\u7684\u7a7a\u95f4\u4e2d\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u6280\u672f\u7684\u7f3a\u70b9\u662f\uff0c\u4f60\u4f1a\u56e0\u4e3a\u964d\u4f4e\u539f\u59cb\u53c2\u6570\u7684\u7cbe\u5ea6\u800c\u635f\u5931\u6027\u80fd\u3002\u7cbe\u5ea6\u8d8a\u4f4e\uff0c\u6027\u80fd\u635f\u5931\u5c31\u8d8a\u5927\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6d41\u884c\u7684\u91cf\u5316\u6280\u672f<\/h3>\n\n\n\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u8bd5\u7740\u7406\u89e3\u5982\u4f55\u4ee5 32 \u4f4d\u6d6e\u70b9\u8868\u793a\u5f62\u5f0f\u5b58\u50a8\u6570\u5b57\uff0c\u5e76\u4e86\u89e3\u5982\u4f55\u5c06\u5b83\u4eec\u91cf\u5316\u4e3a 16 \u4f4d\u6d6e\u70b9\u8868\u793a\u5f62\u5f0f\u3002\u5728\u5404\u79cd\u6570\u636e\u7c7b\u578b\u4e2d\uff0c\u6d6e\u70b9\u6570\u4e3b\u8981\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\uff0c\u56e0\u4e3a\u5b83\u4eec\u80fd\u591f\u4ee5\u9ad8\u7cbe\u5ea6\u8868\u793a\u5404\u79cd\u503c\u3002\u901a\u5e38\uff0c\u6d6e\u70b9\u6570\u4f7f\u7528<em>n<\/em>\u4f4d\u6765\u5b58\u50a8\u6570\u503c\u3002\u8fd9<em>n<\/em>\u4f4d\u8fdb\u4e00\u6b65\u5212\u5206\u4e3a\u4e09\u4e2a\u4e0d\u540c\u7684\u90e8\u5206\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u7b26\u53f7\uff1a\u6b64\u4f4d\u8868\u793a\u6570\u5b57\u7684\u6b63\u6570\u6216\u8d1f\u6570\u3002\u5b83\u4f7f\u7528\u4e00\u4e2a\u4f4d\uff0c\u5176\u4e2d 0 \u8868\u793a\u6b63\u6570\uff0c1 \u8868\u793a\u8d1f\u6570\u3002<\/li>\n\n\n\n<li>\u6307\u6570\uff1a\u6307\u6570\u662f\u4e00\u6bb5\u8868\u793a\u5e95\u6570\uff08\u4e8c\u8fdb\u5236\u8868\u793a\u4e2d\u901a\u5e38\u4e3a 2\uff09\u7684\u5e42\u7684\u4f4d\u3002\u6307\u6570\u4e5f\u53ef\u4ee5\u662f\u6b63\u6570\u6216\u8d1f\u6570\uff0c\u4ece\u800c\u4f7f\u6570\u5b57\u53ef\u4ee5\u8868\u793a\u975e\u5e38\u5927\u6216\u975e\u5e38\u5c0f\u7684\u503c\u3002<\/li>\n\n\n\n<li>\u6709\u6548\u4f4d\/\u5c3e\u6570\uff1a\u5269\u4f59\u7684\u4f4d\u7528\u4e8e\u5b58\u50a8\u6709\u6548\u4f4d\uff0c\u4e5f\u79f0\u4e3a\u5c3e\u6570\u3002\u8fd9\u4ee3\u8868\u6570\u5b57\u7684\u6709\u6548\u6570\u5b57\u3002\u6570\u5b57\u7684\u7cbe\u5ea6\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u6709\u6548\u4f4d\u7684\u957f\u5ea6\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u79cd\u8bbe\u8ba1\u5141\u8bb8\u6d6e\u70b9\u6570\u4ee5\u4e0d\u540c\u7684\u7cbe\u5ea6\u8986\u76d6\u5e7f\u6cdb\u7684\u503c\u3002\u8fd9\u79cd\u8868\u793a\u4f7f\u7528\u7684\u516c\u5f0f\u662f\uff1a<\/p>\n\n\n\n<p>(\u22121)<em>\u7b26\u53f7<\/em>\u2217<em>\u5e95\u6570<\/em><em>\u6307\u6570<\/em>\u2217<em>\u5c3e\u6570<\/em><\/p>\n\n\n\n<p>\u4e3a\u4e86\u66f4\u597d\u5730\u7406\u89e3\u8fd9\u4e00\u70b9\uff0c\u8ba9\u6211\u4eec\u770b\u770b\u6df1\u5ea6\u5b66\u4e60\u4e2d\u6700\u5e38\u7528\u7684\u4e00\u4e9b\u6570\u636e\u7c7b\u578b\uff0c\u770b\u770b\u5982\u4f55\u5c06 float32 \u6570\u5b57\u8f6c\u6362\u4e3a float16 \u548c bfloat16 \u6570\u5b57\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>FP32\u4f7f\u7528 32 \u4f4d\u6765\u8868\u793a\u6570\u5b57\uff1a1 \u4f4d\u8868\u793a\u7b26\u53f7\uff0c8 \u4f4d\u8868\u793a\u6307\u6570\uff0c\u5176\u4f59 23 \u4f4d\u8868\u793a\u6709\u6548\u6570\u5b57\u3002\u867d\u7136\u5b83\u63d0\u4f9b\u4e86\u9ad8\u7cbe\u5ea6\uff0c\u4f46 FP32 \u7684\u7f3a\u70b9\u662f\u8ba1\u7b97\u548c\u5185\u5b58\u5360\u7528\u91cf\u5927\u3002\u8003\u8651\u4ee5\u4e0b 2.71828 \u7684 FP32 \u8868\u793a\uff1a<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGKJkeya3Gx1s2tHKib2K14qH6Ga8EWzjgwoUr3mx1GYCbhV4Bh5RGUVw\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"\u91cf\u5316_fp32drawio \u2013 LearnOpenCV\u00a0\"\/><\/figure>\n\n\n\n<p>\uff08-1\uff090&nbsp;\u2217&nbsp;2128-127&nbsp;\u22171.35914003=2.71828007<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>FP16\u4f7f\u7528 16 \u4f4d\u6765\u5b58\u50a8\u6570\u5b57\uff1a1 \u7528\u4e8e\u7b26\u53f7\uff0c5 \u7528\u4e8e\u6307\u6570\uff0c10 \u7528\u4e8e\u5c3e\u6570\u3002\u867d\u7136\u8fd9\u4f7f\u5176\u66f4\u8282\u7701\u5185\u5b58\u5e76\u52a0\u5feb\u4e86\u8ba1\u7b97\u901f\u5ea6\uff0c\u4f46\u8303\u56f4\u548c\u7cbe\u5ea6\u7684\u964d\u4f4e\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6570\u503c\u4e0d\u7a33\u5b9a\uff0c\u4ece\u800c\u53ef\u80fd\u5f71\u54cd\u6a21\u578b\u51c6\u786e\u6027\u3002\u67e5\u770b 2.71828 \u7684\u91cf\u5316 FP16 \u8868\u793a\uff1a<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGSkFvpgB5VeUgPSMITWC6El1MufOl4hFm5KTe0fyV01KQI7wMjaVfHw\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"\u91cf\u5316_fp16drawio \u2013 LearnOpenCV\u00a0\"\/><\/figure>\n\n\n\n<p>\uff08-1\uff090&nbsp;\u2217&nbsp;2128-127&nbsp;\u22171.358398=2.716796<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>BF16 \u4e5f\u662f16 \u4f4d\u683c\u5f0f\uff0c\u4f46\u5176\u4e2d 1 \u4f4d\u8868\u793a\u7b26\u53f7\uff0c8 \u4f4d\u8868\u793a\u6307\u6570\uff0c7 \u4f4d\u8868\u793a\u6709\u6548\u6570\u5b57\u3002\u4e0e FP16 \u76f8\u6bd4\uff0cBF16 \u6269\u5927\u4e86\u53ef\u8868\u793a\u8303\u56f4\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u4e0b\u6ea2\u548c\u4e0a\u6ea2\u98ce\u9669\u3002\u5c3d\u7ba1\u7531\u4e8e\u6709\u6548\u4f4d\u8f83\u5c11\u5bfc\u81f4\u7cbe\u5ea6\u964d\u4f4e\uff0c\u4f46 BF16 \u901a\u5e38\u4e0d\u4f1a\u663e\u8457\u5f71\u54cd\u6a21\u578b\u6027\u80fd\uff0c\u5e76\u4e14\u5bf9\u4e8e\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u6765\u8bf4\u662f\u4e00\u79cd\u6709\u7528\u7684\u6298\u8877\u65b9\u6848\u3002\u67e5\u770b 2.71828 \u7684\u91cf\u5316\u6d6e\u70b9 16 \u8868\u793a\uff1a<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/mmbiz.qpic.cn\/mmbiz_png\/qt08A648NlNh7Vz0PysIMGlfmHUZpUBGepicWneYcLEOlIyqIrScEkficQ4JRWcEYibNFMJaNHMtn1ffaicGuQs94g\/640?wx_fmt=png&amp;from=appmsg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1\" alt=\"\u56fe\u7247\" title=\"\u91cf\u5316_bf16drawio \u2013 LearnOpenCV\u00a0\"\/><\/figure>\n\n\n\n<p>\uff08-1\uff090&nbsp;\u2217&nbsp;2128-127&nbsp;\u22171.3515625=2.703125<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u8bba<\/h2>\n\n\n\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 Transformer \u67b6\u6784\uff0c\u5e76\u76f4\u89c2\u5730\u4e86\u89e3\u4e86 Transformer \u67b6\u6784\u5982\u4f55\u7406\u89e3\u81ea\u7136\u8bed\u8a00\u3002\u7136\u540e\uff0c\u6211\u4eec\u63a2\u7d22\u4e86\u51e0\u79cd\u5fae\u8c03\u6280\u672f\uff0c\u4f8b\u5982\u5b8c\u5168\u5fae\u8c03\u548c\u53c2\u6570\u9ad8\u6548\u5fae\u8c03 (PEFT)\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u6709\u51e0\u79cd PEFT \u6280\u672f\u53ef\u7528\u4e8e\u4f7f\u7528\u5355\u4e2a GPU \u9488\u5bf9\u60a8\u7684\u72ec\u7279\u6570\u636e\u5bf9 LLM \u8fdb\u884c\u5fae\u8c03\u3002\u6211\u4eec\u8fd8\u4ecb\u7ecd\u4e86\u91cf\u5316\u7684\u6982\u5ff5\uff0c\u5b83\u4f7f\u6211\u4eec\u80fd\u591f\u51cf\u5c11\u6a21\u578b\u7684\u5185\u5b58\u5360\u7528\u5e76\u52a0\u901f\u63a8\u7406\u8fc7\u7a0b\u3002\u5728\u672c\u6587\u7684\u6700\u540e\uff0c\u6211\u4eec\u8ba8\u8bba\u4e86\u4e00\u4e9b\u91cd\u8981\u7684\u8bc4\u4f30\u6307\u6807\uff0c\u8fd9\u4e9b\u6307\u6807\u53ef\u4f5c\u4e3a LLM \u7cfb\u7edf\u5730\u5b66\u4e60\u7ed9\u5b9a\u4efb\u52a1\u7684\u6307\u5357\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u662f\u201c\u638c\u63e1 LLM\u201d\u7cfb\u5217\u7684\u7b2c\u4e00\u90e8\u5206\uff0c\u5728\u8fd9\u91cc\u6211\u4eec\u5c06\u8ba8\u8bba\u4ec0\u4e48\u662f LLM\u3001\u5b83\u4eec\u7684\u7528\u4f8b\u4ee5\u53ca\u5b83\u4eec\u6267\u884c\u7684\u4efb\u52a1\u3002\u6211\u4eec\u8fd8\u5c06\u89e3 [&hellip;]<\/p>\n","protected":false},"author":9,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[29],"tags":[40],"class_list":["post-781","post","type-post","status-publish","format-standard","hentry","category-transformer","tag-40"],"blocksy_meta":"","_links":{"self":[{"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/posts\/781","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/comments?post=781"}],"version-history":[{"count":1,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/posts\/781\/revisions"}],"predecessor-version":[{"id":782,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/posts\/781\/revisions\/782"}],"wp:attachment":[{"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/media?parent=781"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/categories?post=781"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/ai.gitpp.com\/index.php\/wp-json\/wp\/v2\/tags?post=781"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}