{"id":15880,"date":"2026-06-10T07:09:41","date_gmt":"2026-06-09T23:09:41","guid":{"rendered":"https:\/\/ab123.xyz\/?p=15880"},"modified":"2026-06-10T07:09:41","modified_gmt":"2026-06-09T23:09:41","slug":"deepseek-r1-%e5%be%ae%e8%b0%83%ef%bc%9alora-%e4%b8%8e-qlora-%e5%ae%9e%e6%88%98","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=15880","title":{"rendered":"DeepSeek-R1 \u5fae\u8c03\uff1aLoRA \u4e0e QLoRA \u5b9e\u6218"},"content":{"rendered":"<p>\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5feb\u901f\u8fed\u4ee3\u7684\u5f53\u4e0b\uff0cDeepSeek-R1 \u51ed\u501f\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u5907\u53d7\u5173\u6ce8\u3002\u7136\u800c\uff0c\u76f4\u63a5\u5168\u53c2\u6570\u5fae\u8c03\u4e0d\u4ec5\u6d88\u8017\u5de8\u5927\u7b97\u529b\uff0c\u8fd8\u5bb9\u6613\u5bfc\u81f4\u707e\u96be\u6027\u9057\u5fd8\u3002LoRA\uff08Low-Rank Adaptation\uff09\u4e0e QLoRA\uff08Quantized Low-Rank Adaptation\uff09\u4e24\u79cd\u8f7b\u91cf\u5316\u5fae\u8c03\u6280\u672f\u5e94\u8fd0\u800c\u751f\uff0c\u4e3a\u5f00\u53d1\u8005\u63d0\u4f9b\u4e86\u9ad8\u6548\u3001\u4f4e\u6210\u672c\u7684\u5b9e\u8df5\u8def\u5f84\u3002\u672c\u6587\u5e26\u4f60\u6df1\u5165\u89e3\u6790\u8fd9\u4e24\u79cd\u65b9\u6cd5\u7684\u539f\u7406\u4e0e\u5b9e\u6218\u6280\u5de7\uff0c\u5e76\u9644\u4e0a\u5b98\u65b9\u8d44\u6e90\u5165\u53e3\u3002<\/p>\n<p>\u5b98\u65b9\u5de5\u5177\u4e0e\u6559\u7a0b\u5165\u53e3\uff1a<a href=\"https:\/\/www.deepseek.com\" target=\"_blank\">DeepSeek \u5b98\u65b9\u7f51\u7ad9<\/a>\uff0c\u83b7\u53d6\u6700\u65b0\u6a21\u578b\u6743\u91cd\u4e0e\u5fae\u8c03\u811a\u672c\u3002<\/p>\n<h2>LoRA \u5fae\u8c03\uff1a\u53c2\u6570\u9ad8\u6548\u7684\u6838\u5fc3\u673a\u5236<\/h2>\n<h3>\u539f\u7406\u4e0e\u4f18\u52bf<\/h3>\n<p>LoRA \u901a\u8fc7\u51bb\u7ed3\u539f\u6a21\u578b\u53c2\u6570\uff0c\u4ec5\u5bf9\u6bcf\u4e00\u5c42\u6ce8\u5165\u4f4e\u79e9\u77e9\u9635\u8fdb\u884c\u8bad\u7ec3\u3002\u8fd9\u4f7f\u5f97\u53c2\u6570\u91cf\u51cf\u5c11\u81f3\u5168\u91cf\u5fae\u8c03\u7684 0.1% \u4ee5\u4e0b\uff0c\u663e\u5b58\u5360\u7528\u5927\u5e45\u964d\u4f4e\uff0c\u4e14\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347 3 \u500d\u4ee5\u4e0a\u3002\u4ee5 DeepSeek-R1 \u4e3a\u4f8b\uff0c\u4f7f\u7528 LoRA \u5fae\u8c03\u540e\uff0c\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u3001\u6570\u5b66\u63a8\u7406\u7b49\u4e0b\u6e38\u4efb\u52a1\u4e2d\u8868\u73b0\u51e0\u4e4e\u65e0\u635f\uff0c\u9002\u5408\u4e2a\u4eba\u5f00\u53d1\u8005\u6216\u5c0f\u56e2\u961f\u5feb\u901f\u5b9e\u9a8c\u3002<\/p>\n<h3>\u5173\u952e\u914d\u7f6e\u4e0e\u6280\u5de7<\/h3>\n<ul>\n<li>\u79e9\uff08rank\uff09\uff1a\u63a8\u8350\u53d6\u503c 8-16\uff0c\u5e73\u8861\u9002\u914d\u80fd\u529b\u4e0e\u6548\u7387<\/li>\n<li>\u7f29\u653e\u56e0\u5b50\uff08alpha\uff09\uff1a\u901a\u5e38\u4e3a rank \u7684 2 \u500d\uff0c\u907f\u514d\u68af\u5ea6\u7206\u70b8<\/li>\n<li>\u76ee\u6807\u6a21\u5757\uff1a\u4f18\u5148\u9009\u62e9 query \u4e0e value \u77e9\u9635\uff0c\u4e5f\u53ef\u6269\u5c55\u81f3\u5168\u8fde\u63a5\u5c42<\/li>\n<\/ul>\n<h2>QLoRA \u5fae\u8c03\uff1a\u5728\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u8fd0\u884c<\/h2>\n<h3>4 \u4f4d\u91cf\u5316\u4e0e\u53cc\u91cf\u5316\u6280\u672f<\/h3>\n<p>QLoRA \u5728 LoRA \u57fa\u7840\u4e0a\u5f15\u5165 NF4 \u91cf\u5316\u4e0e\u53cc\u91cf\u5316\u6280\u672f\uff0c\u5c06\u6a21\u578b\u6743\u91cd\u538b\u7f29\u81f3 4 \u4f4d\uff0c\u663e\u5b58\u9700\u6c42\u518d\u964d 75%\u3002\u8fd9\u610f\u5473\u7740\u5355\u5f20 RTX 4090\uff0824GB\uff09\u5373\u53ef\u5fae\u8c03 DeepSeek-R1 \u7684 70B \u7248\u672c\u3002\u901a\u8fc7\u5206\u9875\u4f18\u5316\u5668\u548c\u68af\u5ea6\u68c0\u67e5\u70b9\uff0c\u5373\u4f7f\u663e\u5b58\u4e0d\u8db3\u4e5f\u80fd\u6d41\u7545\u8fd0\u884c\u3002<\/p>\n<h3>\u5b9e\u6218\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li>\u91cf\u5316\u7c7b\u578b\uff1aNF4 \u6bd4 FP4 \u66f4\u7a33\u5b9a\uff0c\u63a8\u8350\u7528\u4e8e\u63a8\u7406\u5bc6\u96c6\u578b\u4efb\u52a1<\/li>\n<li>\u9002\u914d\u5668\u5408\u5e76\uff1a\u8bad\u7ec3\u540e\u9700\u5c06 LoRA \u6743\u91cd\u4e0e\u91cf\u5316\u4e3b\u6a21\u578b\u5408\u5e76\uff0c\u518d\u4fdd\u5b58\u4e3a\u5b8c\u6574\u7cbe\u5ea6<\/li>\n<li>\u6570\u636e\u96c6\u8d28\u91cf\uff1a\u4f4e\u6bd4\u7279\u8bad\u7ec3\u5bf9\u566a\u58f0\u654f\u611f\uff0c\u5efa\u8bae\u6e05\u6d17\u6570\u636e\u5e76\u589e\u52a0\u6307\u4ee4\u6a21\u677f<\/li>\n<\/ul>\n<h2>\u5e94\u7528\u573a\u666f\u4e0e\u843d\u5730\u5efa\u8bae<\/h2>\n<p>LoRA \u4e0e QLoRA \u7279\u522b\u9002\u5408\u4ee5\u4e0b\u573a\u666f\uff1a\u5782\u76f4\u9886\u57df\u5ba2\u670d\u673a\u5668\u4eba\uff08\u91d1\u878d\u3001\u533b\u7597\uff09\u3001\u4ee3\u7801\u81ea\u52a8\u8865\u5168\u63d2\u4ef6\u3001\u4e2a\u6027\u5316\u6559\u80b2\u8f85\u5bfc\u3002\u4f7f\u7528\u65f6\u9700\u6ce8\u610f\uff1a\u82e5\u4efb\u52a1\u590d\u6742\u4e14\u6570\u636e\u91cf\u8d85\u8fc7 1 \u4e07\u6761\uff0c\u5efa\u8bae\u5207\u6362\u4e3a\u5168\u53c2\u6570\u5fae\u8c03\uff1b\u82e5\u8ffd\u6c42\u6781\u81f4\u90e8\u7f72\u901f\u5ea6\uff0c\u53ef\u5bfc\u51fa\u4e3a ONNX \u683c\u5f0f\u5e76\u914d\u5408 vLLM \u63a8\u7406\u3002\u5f53\u524d\u5df2\u6709\u793e\u533a\u57fa\u4e8e DeepSeek-R1 \u7684 LoRA \u6743\u91cd\u5728 Hugging Face \u4e0a\u5f00\u6e90\uff0c\u53ef\u76f4\u63a5\u4e0b\u8f7d\u4f7f\u7528\u3002<\/p>\n<h2>\u4ece\u96f6\u5f00\u59cb\u7684\u5fae\u8c03\u6b65\u9aa4<\/h2>\n<h3>\u73af\u5883\u51c6\u5907<\/h3>\n<p>\u63a8\u8350 Python 3.10 + PyTorch 2.1\uff0c\u5b89\u88c5 transformers \u4e0e peft \u5e93\u3002\u4ece DeepSeek \u5b98\u65b9\u4ed3\u5e93\u4e0b\u8f7d\u57fa\u5ea7\u6a21\u578b\uff0c\u4f7f\u7528 bitsandbytes \u542f\u7528 4 \u4f4d\u52a0\u8f7d\u3002<\/p>\n<h3>\u8bad\u7ec3\u811a\u672c\u6838\u5fc3\u4ee3\u7801<\/h3>\n<p>\u5b9a\u4e49 LoRA \u914d\u7f6e\u540e\uff0c\u91c7\u7528 Hugging Face Trainer \u8fdb\u884c\u591a\u5361\u8bad\u7ec3\u3002\u5173\u952e\u8d85\u53c2\u6570\uff1a\u5b66\u4e60\u7387 2e-4\uff0c\u6279\u6b21\u5927\u5c0f 4\uff0c\u68af\u5ea6\u7d2f\u79ef 8 \u6b65\u3002\u8bad\u7ec3\u7ea6 3 \u5c0f\u65f6\u5373\u53ef\u5728 48GB \u663e\u5b58\u4e0a\u5b8c\u6210 1 \u4e07\u6761\u6307\u4ee4\u7684\u5fae\u8c03\u3002<\/p>\n<p>\u66f4\u591a\u5b9e\u6218\u793a\u4f8b\u4e0e\u793e\u533a\u8ba8\u8bba\uff0c\u8bf7\u8bbf\u95ee <a href=\"https:\/\/www.deepseek.com\" target=\"_blank\">DeepSeek \u5b98\u65b9\u7f51\u7ad9<\/a> \u7684\u300c\u5f00\u53d1\u8005\u6587\u6863\u300d\u677f\u5757\u3002\u6301\u7eed\u5173\u6ce8\u6700\u65b0\u7248\u672c\u66f4\u65b0\uff0c\u53ef\u83b7\u53d6\u9488\u5bf9 LoRA\/QLoRA \u7684\u4f18\u5316\u7b97\u5b50\u652f\u6301\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5feb\u901f\u8fed\u4ee3\u7684\u5f53\u4e0b\uff0cDeepSeek-R1 \u51ed\u501f\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u5907\u53d7\u5173\u6ce8\u3002\u7136\u800c\uff0c\u76f4\u63a5\u5168\u53c2\u6570\u5fae\u8c03\u4e0d\u4ec5\u6d88\u8017 [&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":[10877,9267,19096,19097,18821],"class_list":["post-15880","post","type-post","status-publish","format-standard","hentry","category-4","tag-deepseek-r1","tag-lora","tag-qlora","tag-19097","tag-18821"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/15880","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=15880"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/15880\/revisions"}],"predecessor-version":[{"id":15881,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/15880\/revisions\/15881"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=15880"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=15880"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=15880"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}