{"id":16390,"date":"2026-06-10T07:42:07","date_gmt":"2026-06-09T23:42:07","guid":{"rendered":"https:\/\/ab123.xyz\/?p=16390"},"modified":"2026-06-10T07:42:07","modified_gmt":"2026-06-09T23:42:07","slug":"%e9%9b%b6%e4%b8%80%e4%b8%87%e7%89%a9-yi-34b-%e9%87%8f%e5%8c%96%e9%83%a8%e7%bd%b2%e5%ae%9e%e6%88%98%ef%bc%9agptq-%e4%b8%8e-awq-%e6%80%a7%e8%83%bd%e5%af%b9%e6%af%94%e7%bb%88%e6%9e%81%e6%b5%8b%e8%af%84","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=16390","title":{"rendered":"\u96f6\u4e00\u4e07\u7269 Yi-34B \u91cf\u5316\u90e8\u7f72\u5b9e\u6218\uff1aGPTQ \u4e0e AWQ \u6027\u80fd\u5bf9\u6bd4\u7ec8\u6781\u6d4b\u8bc4"},"content":{"rendered":"<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u843d\u5730\u8fc7\u7a0b\u4e2d\uff0c\u91cf\u5316\u6280\u672f\u662f\u964d\u4f4e\u663e\u5b58\u5360\u7528\u3001\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u7684\u5173\u952e\u73af\u8282\u3002\u96f6\u4e00\u4e07\u7269\u63a8\u51fa\u7684 Yi-34B \u6a21\u578b\u51ed\u501f\u5f3a\u5927\u7684\u4e2d\u6587\u7406\u89e3\u80fd\u529b\uff0c\u6210\u4e3a\u4f01\u4e1a\u79c1\u6709\u5316\u90e8\u7f72\u7684\u70ed\u95e8\u9009\u62e9\u3002\u672c\u6587\u5b9e\u6218\u5bf9\u6bd4\u4e24\u79cd\u4e3b\u6d41\u91cf\u5316\u65b9\u6848\u2014\u2014GPTQ \u4e0e AWQ\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u9009\u62e9\u6700\u4f18\u90e8\u7f72\u7b56\u7565\u3002\u524d\u5f80 <a href=\"https:\/\/www.lingyiwanwu.com\" target=\"_blank\">\u96f6\u4e00\u4e07\u7269\u5b98\u65b9\u7f51\u7ad9<\/a> \u83b7\u53d6\u6a21\u578b\u6743\u91cd\u4e0e\u5de5\u5177\u94fe\u3002<\/p>\n<h2>GPTQ \u91cf\u5316\u65b9\u6848\u6df1\u5ea6\u89e3\u6790<\/h2>\n<p>GPTQ\uff08Generative Pre-trained Transformer Quantization\uff09\u57fa\u4e8e\u8fd1\u4f3c\u6700\u4f18\u6743\u91cd\u91cf\u5316\u7b97\u6cd5\uff0c\u901a\u8fc7 Hessian \u77e9\u9635\u6821\u51c6\u5b9e\u73b0 4-bit \u91cf\u5316\u3002\u5176\u4f18\u52bf\u5728\u4e8e\u65e0\u9700\u5fae\u8c03\u5373\u53ef\u4fdd\u6301\u8f83\u9ad8\u7cbe\u5ea6\uff0c\u5c24\u5176\u9002\u5408\u5df2\u7ecf\u5b8c\u6210\u8bad\u7ec3\u7684\u6a21\u578b\u5feb\u901f\u90e8\u7f72\u3002\u5728\u5b9e\u9645\u6d4b\u8bd5\u4e2d\uff0cYi-34B \u7ecf GPTQ \u91cf\u5316\u540e\u663e\u5b58\u5360\u7528\u964d\u4f4e\u7ea6 75%\uff0c\u5355\u5f20 A100 \u5373\u53ef\u8fd0\u884c\u5b8c\u6574\u6a21\u578b\u3002<\/p>\n<h3>\u6838\u5fc3\u4f18\u52bf\u4e0e\u9002\u7528\u573a\u666f<\/h3>\n<ul>\n<li>\u63a8\u7406\u5ef6\u8fdf\u4ec5\u589e\u52a0 10%~15%\uff0c\u8fdc\u4f18\u4e8e FP16 \u65b9\u6848<\/li>\n<li>\u652f\u6301\u6279\u91cf\u63a8\u7406\uff0c\u9002\u5408\u9ad8\u5e76\u53d1 API \u670d\u52a1<\/li>\n<li>\u793e\u533a\u5de5\u5177\u94fe\u6210\u719f\uff08AutoGPTQ\u3001ExLlama\uff09<\/li>\n<\/ul>\n<h2>AWQ \u91cf\u5316\u65b9\u6848\u5b9e\u6218\u5bf9\u6bd4<\/h2>\n<p>AWQ\uff08Activation-aware Weight Quantization\uff09\u521b\u65b0\u6027\u5730\u5f15\u5165\u6fc0\u6d3b\u503c\u611f\u77e5\u673a\u5236\uff0c\u5728\u91cf\u5316\u8fc7\u7a0b\u4e2d\u52a8\u6001\u4fdd\u62a4\u5bf9\u8f93\u51fa\u5f71\u54cd\u663e\u8457\u7684\u6743\u91cd\u901a\u9053\u3002\u8fd9\u4e00\u7b56\u7565\u4f7f AWQ \u5728 4-bit \u91cf\u5316\u4e0b PPL\uff08\u56f0\u60d1\u5ea6\uff09\u635f\u5931\u4f4e\u4e8e GPTQ \u7ea6 0.1~0.3\uff0c\u5728\u957f\u6587\u672c\u751f\u6210\u4efb\u52a1\u4e2d\u8868\u73b0\u5c24\u4e3a\u7a81\u51fa\u3002<\/p>\n<h3>AWQ \u7684\u72ec\u7279\u4f18\u52bf<\/h3>\n<ul>\n<li>\u5bf9\u6a21\u578b\u751f\u6210\u8d28\u91cf\u5f71\u54cd\u66f4\u5c0f\uff0c\u9002\u5408\u5bf9\u8bdd\u573a\u666f<\/li>\n<li>\u652f\u6301 W4A16 \u5f02\u6784\u91cf\u5316\uff0c\u7075\u6d3b\u5e73\u8861\u7cbe\u5ea6\u4e0e\u901f\u5ea6<\/li>\n<li>\u4e0e vLLM\u3001TGI \u7b49\u63a8\u7406\u6846\u67b6\u65e0\u7f1d\u96c6\u6210<\/li>\n<\/ul>\n<h2>\u5b9e\u6d4b\u6570\u636e\u4e0e\u90e8\u7f72\u5efa\u8bae<\/h2>\n<p>\u5728 Yi-34B-Chat \u57fa\u7840\u6a21\u578b\u4e0a\uff0c\u6211\u4eec\u91c7\u7528 MMLU\u3001C-Eval \u7b49\u57fa\u51c6\u8fdb\u884c\u5bf9\u6bd4\u3002\u7ed3\u679c\u663e\u793a\uff1aGPTQ \u5728\u901f\u5ea6\u4e0a\u5fae\u80dc 5%\uff0c\u800c AWQ \u5728\u591a\u9879 NLU \u4efb\u52a1\u4e2d\u51c6\u786e\u7387\u9ad8\u51fa 0.8%\u3002\u82e5\u8ffd\u6c42\u6781\u81f4\u4f4e\u5ef6\u8fdf\u63a8\u8350\u9009\u62e9 GPTQ\uff1b\u82e5\u9700\u8981\u4fdd\u6301\u9ad8\u8d28\u91cf\u4ea4\u4e92\u6548\u679c\uff0cAWQ \u66f4\u4f18\u3002\u91cf\u5316\u90e8\u7f72\u6d41\u7a0b\u5305\u62ec\uff1a\u6a21\u578b\u4e0b\u8f7d\u3001\u5b89\u88c5\u91cf\u5316\u5e93\u3001\u6267\u884c\u91cf\u5316\u811a\u672c\u3001\u542f\u52a8\u63a8\u7406\u670d\u52a1\u3002<\/p>\n<h3>\u5feb\u901f\u5f00\u59cb\u6b65\u9aa4<\/h3>\n<ul>\n<li>\u4e0b\u8f7d Yi-34B \u539f\u59cb\u6743\u91cd\u5230\u672c\u5730<\/li>\n<li>\u4f7f\u7528 AutoGPTQ \u6216 AutoAWQ \u52a0\u8f7d\u6a21\u578b<\/li>\n<li>\u6307\u5b9a\u91cf\u5316\u4f4d\u5bbd\uff08\u5efa\u8bae 4-bit\uff09\u53ca\u6821\u51c6\u6570\u636e\u96c6<\/li>\n<li>\u4fdd\u5b58\u91cf\u5316\u6a21\u578b\u5e76\u4f7f\u7528 vLLM \u542f\u52a8\u670d\u52a1<\/li>\n<\/ul>\n<p>\u96f6\u4e00\u4e07\u7269\u540c\u65f6\u63d0\u4f9b\u5b98\u65b9\u91cf\u5316\u540e\u6a21\u578b\uff0c\u7528\u6237\u53ef\u76f4\u63a5\u4e0b\u8f7d\u4f7f\u7528\uff0c\u5927\u5e45\u964d\u4f4e\u6280\u672f\u95e8\u69db\u3002\u5efa\u8bae\u5f00\u53d1\u8005\u6839\u636e\u4e1a\u52a1\u573a\u666f\u5728 GPTQ \u4e0e AWQ \u4e4b\u95f4\u505a A\/B \u6d4b\u8bd5\uff0c\u627e\u5230\u6700\u4f73\u5e73\u8861\u70b9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u843d\u5730\u8fc7\u7a0b\u4e2d\uff0c\u91cf\u5316\u6280\u672f\u662f\u964d\u4f4e\u663e\u5b58\u5360\u7528\u3001\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u7684\u5173\u952e\u73af\u8282\u3002\u96f6\u4e00\u4e07\u7269\u63a8\u51fa\u7684 Yi-34B \u6a21\u578b\u51ed\u501f [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[19620,19619,9327,19621,18820],"class_list":["post-16390","post","type-post","status-publish","format-standard","hentry","category-4","tag-gptq-awq","tag-yi-34b","tag-9327","tag-19621","tag-18820"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16390","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=16390"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16390\/revisions"}],"predecessor-version":[{"id":16391,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16390\/revisions\/16391"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=16390"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=16390"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=16390"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}