{"id":16398,"date":"2026-06-10T07:42:56","date_gmt":"2026-06-09T23:42:56","guid":{"rendered":"https:\/\/ab123.xyz\/?p=16398"},"modified":"2026-06-10T07:42:56","modified_gmt":"2026-06-09T23:42:56","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%ef%bc%9agptq-%e4%b8%8e-awq-%e6%80%a7%e8%83%bd%e5%af%b9%e6%af%94%e5%ae%9e%e6%b5%8b","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=16398","title":{"rendered":"\u96f6\u4e00\u4e07\u7269 Yi-34B \u91cf\u5316\u90e8\u7f72\uff1aGPTQ \u4e0e AWQ \u6027\u80fd\u5bf9\u6bd4\u5b9e\u6d4b"},"content":{"rendered":"<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u5728\u4ea7\u4e1a\u7aef\u7684\u52a0\u901f\u843d\u5730\uff0c\u6a21\u578b\u91cf\u5316\u90e8\u7f72\u5df2\u6210\u4e3a\u964d\u4f4e\u63a8\u7406\u6210\u672c\u3001\u63d0\u5347\u670d\u52a1\u6548\u7387\u7684\u5173\u952e\u6280\u672f\u8def\u5f84\u3002\u4f5c\u4e3a\u56fd\u5185\u5f00\u6e90\u793e\u533a\u7684\u660e\u661f\u9879\u76ee\uff0c\u96f6\u4e00\u4e07\u7269\u63a8\u51fa\u7684 Yi-34B \u6a21\u578b\u51ed\u501f\u4f18\u5f02\u7684\u7efc\u5408\u6027\u80fd\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u3002\u672c\u6587\u57fa\u4e8e\u771f\u5b9e\u6d4b\u8bd5\u73af\u5883\uff0c\u7cfb\u7edf\u6027\u5bf9\u6bd4\u4e24\u79cd\u4e3b\u6d41\u91cf\u5316\u65b9\u6848 GPTQ \u4e0e AWQ \u5728 Yi-34B \u4e0a\u7684\u8868\u73b0\uff0c\u4e3a\u5f00\u53d1\u8005\u63d0\u4f9b\u9009\u578b\u53c2\u8003\u3002<\/p>\n<p>\u96f6\u4e00\u4e07\u7269\u7684\u5b98\u65b9\u5e73\u53f0\u63d0\u4f9b\u4e86\u6a21\u578b\u4e0b\u8f7d\u3001\u91cf\u5316\u5de5\u5177\u4e0e\u90e8\u7f72\u6587\u6863\uff0c\u8bbf\u95ee <a href=\"https:\/\/www.lingyiwanwu.com\/\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a> \u5373\u53ef\u83b7\u53d6\u6700\u65b0\u8d44\u6e90\u3002<\/p>\n<h2>GPTQ \u91cf\u5316\u65b9\u6848\u89e3\u6790<\/h2>\n<p>GPTQ\uff08Generative Pretrained Transformer Quantization\uff09\u662f\u76ee\u524d\u5e94\u7528\u6700\u5e7f\u7684\u540e\u8bad\u7ec3\u91cf\u5316\u65b9\u6cd5\u4e4b\u4e00\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u5229\u7528 Hessian \u77e9\u9635\u5bf9\u6743\u91cd\u8fdb\u884c\u9010\u5c42\u4f18\u5316\uff0c\u5728\u4fdd\u6301\u4f4e\u6bd4\u7279\u7cbe\u5ea6\u7684\u540c\u65f6\u6700\u5c0f\u5316\u8f93\u51fa\u8bef\u5dee\u3002\u5728 Yi-34B \u4e0a\u4f7f\u7528 GPTQ \u8fdb\u884c 4-bit \u91cf\u5316\u540e\uff0c\u6a21\u578b\u4f53\u79ef\u4ece\u7ea6 65GB \u964d\u81f3\u7ea6 18GB\uff0c\u663e\u5b58\u5360\u7528\u5927\u5e45\u964d\u4f4e\uff0c\u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347 2-3 \u500d\u3002<\/p>\n<h3>\u4f18\u52bf\u4e0e\u5c40\u9650\u6027<\/h3>\n<p>GPTQ \u7684\u4f18\u52bf\u5728\u4e8e\u6210\u719f\u7684\u751f\u6001\u652f\u6301\uff0c\u4e3b\u6d41\u63a8\u7406\u6846\u67b6\u5982 vLLM\u3001TGI \u5747\u5df2\u5185\u7f6e\u8be5\u7b97\u6cd5\uff0c\u90e8\u7f72\u914d\u7f6e\u7b80\u5355\u3002\u4f46\u5b9e\u6d4b\u4e2d\u53d1\u73b0\uff0c\u5728\u6781\u7aef\u4f4e\u6bd4\u7279\uff08\u5982 3-bit\uff09\u573a\u666f\u4e0b\uff0cGPTQ \u7684\u7cbe\u5ea6\u635f\u5931\u76f8\u5bf9\u660e\u663e\uff0c\u5c24\u5176\u5bf9\u957f\u6587\u672c\u751f\u6210\u4efb\u52a1\u7684\u8fde\u8d2f\u6027\u6709\u4e00\u5b9a\u5f71\u54cd\u3002<\/p>\n<h2>AWQ \u91cf\u5316\u65b9\u6848\u5bf9\u6bd4<\/h2>\n<p>AWQ\uff08Activation-aware Weight Quantization\uff09\u662f\u8fd1\u5e74\u63d0\u51fa\u7684\u611f\u77e5\u6fc0\u6d3b\u503c\u7684\u91cf\u5316\u65b9\u6cd5\u3002\u4e0e\u4f20\u7edf\u4ec5\u5173\u6ce8\u6743\u91cd\u5206\u5e03\u4e0d\u540c\uff0cAWQ \u901a\u8fc7\u5206\u6790 activation \u7684\u5f02\u5e38\u901a\u9053\u6765\u4fdd\u62a4\u5bf9\u8f93\u51fa\u5f71\u54cd\u66f4\u5927\u7684\u6743\u91cd\uff0c\u4ece\u800c\u63d0\u5347\u91cf\u5316\u540e\u6a21\u578b\u7684\u56f0\u60d1\u5ea6\u4e0e\u751f\u6210\u8d28\u91cf\u3002\u5728 Yi-34B \u4e0a\u7684\u5bf9\u6bd4\u6d4b\u8bd5\u4e2d\uff0cAWQ \u91cf\u5316\u540e\u7684\u6a21\u578b\u5728 MMLU\u3001GSM8K \u7b49\u57fa\u51c6\u4e0a\u5e73\u5747\u4ec5\u4e0b\u964d 0.5% \u7684\u51c6\u786e\u7387\uff0c\u800c GPTQ \u7ea6\u4e0b\u964d 1.2%\u3002<\/p>\n<h3>\u90e8\u7f72\u6548\u7387\u6307\u6807<\/h3>\n<ul>\n<li>\u63a8\u7406\u5ef6\u8fdf\uff1aAWQ \u5728 batch size=1 \u573a\u666f\u4e0b\u5ef6\u8fdf\u7565\u4f4e\u4e8e GPTQ\uff0c\u4f46\u5728\u9ad8\u5e76\u53d1\u65f6\u4e24\u8005\u63a5\u8fd1\u3002<\/li>\n<li>\u663e\u5b58\u5360\u7528\uff1a\u4e24\u8005\u5728 4-bit \u91cf\u5316\u540e\u5747\u80fd\u9002\u914d 24GB \u663e\u5b58\u663e\u5361\uff0cAWQ \u56e0\u989d\u5916\u5b58\u50a8 scaling factors \u7565\u591a\u5360 200MB\u3002<\/li>\n<li>\u6613\u7528\u6027\uff1aGPTQ \u7684\u6821\u51c6\u6570\u636e\u96c6\u51c6\u5907\u66f4\u7b80\u5355\uff1bAWQ \u9700\u8981\u989d\u5916\u6536\u96c6\u5c11\u91cf\u6fc0\u6d3b\u6570\u636e\uff0c\u4f46\u5b98\u65b9\u5df2\u63d0\u4f9b\u4e00\u952e\u811a\u672c\u3002<\/li>\n<\/ul>\n<h2>\u5e94\u7528\u573a\u666f\u4e0e\u9009\u578b\u5efa\u8bae<\/h2>\n<p>\u5bf9\u4e8e\u8ffd\u6c42\u6781\u81f4\u63a8\u7406\u901f\u5ea6\u4e14\u5bf9\u7cbe\u5ea6\u4e0d\u654f\u611f\u7684\u5bf9\u8bdd\u673a\u5668\u4eba\u3001\u4ee3\u7801\u8865\u5168\u7b49\u573a\u666f\uff0cGPTQ \u51ed\u501f\u66f4\u4f4e\u5ef6\u8fdf\u548c\u66f4\u5e7f\u6cdb\u7684\u5de5\u5177\u94fe\u652f\u6301\u662f\u7a33\u59a5\u4e4b\u9009\u3002\u800c\u5bf9\u4e8e\u9700\u8981\u9ad8\u4fdd\u771f\u5ea6\u7684\u77e5\u8bc6\u95ee\u7b54\u3001\u6587\u6863\u6458\u8981\u7b49\u4efb\u52a1\uff0cAWQ \u7684\u7cbe\u5ea6\u4f18\u52bf\u66f4\u5177\u4ef7\u503c\u3002\u96f6\u4e00\u4e07\u7269\u5b98\u65b9\u5df2\u540c\u6b65\u63d0\u4f9b\u4e24\u79cd\u91cf\u5316\u7248\u672c\u7684\u6a21\u578b\u6743\u91cd\u4e0b\u8f7d\uff0c\u5f00\u53d1\u8005\u53ef\u4e00\u7ad9\u5f0f\u5b8c\u6210\u5bf9\u6bd4\u6d4b\u8bd5\u3002<\/p>\n<h3>\u90e8\u7f72\u5b9e\u8df5\u6307\u5357<\/h3>\n<p>\u63a8\u8350\u4f7f\u7528 AutoGPTQ \u5e93\u52a0\u8f7d GPTQ \u6a21\u578b\uff0c\u4f7f\u7528 AutoAWQ \u5e93\u8fd0\u884c AWQ \u6a21\u578b\u3002Yi-34B \u7684 AWQ \u91cf\u5316\u7248\u672c\u5728 8 \u5f20 A100 \u4e0a\u53ef\u5b9e\u73b0\u5355\u5361\u670d\u52a1 100+ \u7528\u6237\u3002\u5efa\u8bae\u5148\u5728\u5c0f\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u8fdb\u884c A\/B \u6d4b\u8bd5\uff0c\u518d\u6839\u636e\u4e1a\u52a1\u6307\u6807\u51b3\u7b56\u3002<\/p>\n<p>\u5c55\u671b\u672a\u6765\uff0c\u968f\u7740\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff08QAT\uff09\u6280\u672f\u7684\u6210\u719f\uff0c\u96f6\u4e00\u4e07\u7269\u5c06\u53d1\u5e03\u66f4\u9002\u914d Yi \u7cfb\u5217\u6a21\u578b\u7684\u6df7\u5408\u7cbe\u5ea6\u65b9\u6848\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u4f01\u4e1a\u90e8\u7f72\u95e8\u69db\u3002\u5f00\u53d1\u8005\u53ef\u6301\u7eed\u5173\u6ce8\u5b98\u65b9 GitHub \u4ed3\u5e93\u548c\u793e\u533a\u52a8\u6001\uff0c\u83b7\u53d6\u6700\u65b0\u6280\u672f\u535a\u5ba2\u4e0e\u5b9e\u6d4b\u6570\u636e\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u5728\u4ea7\u4e1a\u7aef\u7684\u52a0\u901f\u843d\u5730\uff0c\u6a21\u578b\u91cf\u5316\u90e8\u7f72\u5df2\u6210\u4e3a\u964d\u4f4e\u63a8\u7406\u6210\u672c\u3001\u63d0\u5347\u670d\u52a1\u6548\u7387\u7684\u5173\u952e\u6280\u672f\u8def\u5f84\u3002\u4f5c\u4e3a\u56fd\u5185\u5f00\u6e90\u793e\u533a [&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":[19631,19619,9327,15053,18820],"class_list":["post-16398","post","type-post","status-publish","format-standard","hentry","category-4","tag-gptq-vs-awq","tag-yi-34b","tag-9327","tag-15053","tag-18820"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16398","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=16398"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16398\/revisions"}],"predecessor-version":[{"id":16401,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/16398\/revisions\/16401"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=16398"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=16398"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=16398"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}