{"id":4104,"date":"2024-07-16T18:35:38","date_gmt":"2024-07-16T10:35:38","guid":{"rendered":"https:\/\/nullthought.net\/?p=4104"},"modified":"2024-10-16T10:45:17","modified_gmt":"2024-10-16T02:45:17","slug":"dora-weight-decomposed-low-rank-adaptation%ef%bc%88%e6%9d%83%e9%87%8d%e5%88%86%e8%a7%a3%e4%bd%8e%e7%a7%a9%e9%80%82%e9%85%8d%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4104","title":{"rendered":"DoRA: Weight-Decomposed Low-Rank Adaptation\uff08\u6743\u91cd\u5206\u89e3\u4f4e\u79e9\u9002\u914d\uff09"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><em><a href=\"https:\/\/arxiv.org\/abs\/2402.09353\" target=\"_blank\" rel=\"noreferrer noopener\">DoRA: Weight-Decomposed Low-Rank Adaptation<\/a><\/em><\/strong>\uff08\u6743\u91cd\u5206\u89e3\u4f4e\u79e9\u9002\u914d\uff09\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u7684\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u65b9\u6cd5\uff08PEFT\uff09\u2014\u2014DoRA\uff0c\u5176\u65e8\u5728\u7f29\u5c0fLoRA\u548c\u5b8c\u5168\u5fae\u8c03\uff08FT\uff09\u4e4b\u95f4\u7684\u51c6\u786e\u6027\u5dee\u8ddd\u3002\u901a\u8fc7\u5c06\u9884\u8bad\u7ec3\u6743\u91cd\u5206\u89e3\u4e3a\u5e45\u5ea6\u548c\u65b9\u5411\u4e24\u4e2a\u90e8\u5206\uff0cDoRA\u4e13\u95e8\u91c7\u7528<a href=\"https:\/\/nullthought.net\/?p=4066\" target=\"_blank\" rel=\"noreferrer noopener\">LoRA<\/a>\u8fdb\u884c\u65b9\u5411\u66f4\u65b0\uff0c\u4ece\u800c\u63d0\u9ad8\u5b66\u4e60\u80fd\u529b\u548c\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u540c\u65f6\u907f\u514d\u989d\u5916\u7684\u63a8\u7406\u6210\u672c\u3002\u5b9e\u9a8c\u8868\u660e\uff0cDoRA\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\uff0c\u5982\u5e38\u8bc6\u63a8\u7406\u548c\u89c6\u89c9\u6307\u4ee4\u5fae\u8c03\uff0c\u5747\u4f18\u4e8eLoRA\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aShih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, Min-Hung Chen\uff0c\u6765\u81eaNVIDIA\u548cHKUST\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u5185\u5bb9\u6982\u8981\u5982\u4e0b\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e00\u3001\u5f15\u8a00<\/h4>\n\n\n\n<p>PEFT\u65b9\u6cd5\u5982LoRA\u56e0\u5176\u9ad8\u6548\u6027\u800c\u5e7f\u53d7\u6b22\u8fce\uff0c\u4f46\u5728\u51c6\u786e\u6027\u4e0a\u5f80\u5f80\u4e0d\u53caFT\u3002DoRA\u901a\u8fc7\u5c06\u6743\u91cd\u5206\u89e3\u4e3a\u5e45\u5ea6\u548c\u65b9\u5411\u7ec4\u4ef6\uff0c\u65e8\u5728\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u6a21\u4eff\u4e86FT\u7684\u5b66\u4e60\u80fd\u529b\uff0c\u8fd8\u63d0\u9ad8\u4e86LoRA\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u540c\u65f6\u6ca1\u6709\u589e\u52a0\u63a8\u7406\u5f00\u9500\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e8c\u3001\u76f8\u5173\u5de5\u4f5c<\/h4>\n\n\n\n<p>PEFT\u65b9\u6cd5\u5206\u4e3a\u4e09\u7c7b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u57fa\u4e8e\u9002\u914d\u5668\u7684\u65b9\u6cd5<\/strong>\uff1a\u5728\u6a21\u578b\u4e2d\u5f15\u5165\u989d\u5916\u7684\u53ef\u8bad\u7ec3\u6a21\u5757\u3002<\/li>\n\n\n\n<li><strong>\u57fa\u4e8e\u63d0\u793a\u7684\u65b9\u6cd5<\/strong>\uff1a\u5728\u8f93\u5165\u4e2d\u6dfb\u52a0\u8f6f\u6807\u8bb0\u3002<\/li>\n\n\n\n<li><strong>\u4f4e\u79e9\u9002\u914d\u65b9\u6cd5\uff08LoRA\u53ca\u5176\u53d8\u79cd\uff09<\/strong>\uff1a\u4f7f\u7528\u4f4e\u79e9\u77e9\u9635\u8fdb\u884c\u6743\u91cd\u66f4\u65b0\u3002<\/li>\n<\/ol>\n\n\n\n<p>DoRA\u5c5e\u4e8e\u7b2c\u4e09\u7c7b\uff0c\u4f46\u901a\u8fc7\u5f15\u5165\u6743\u91cd\u5206\u89e3\u5206\u6790\u6765\u514b\u670dLoRA\u7684\u5c40\u9650\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e09\u3001\u6743\u91cd\u5206\u89e3\u5206\u6790<\/h4>\n\n\n\n<p>\u6743\u91cd\u5206\u89e3\u5206\u6790\u5c06\u6743\u91cd\u91cd\u65b0\u53c2\u6570\u5316\u4e3a\u5e45\u5ea6\u548c\u65b9\u5411\u7ec4\u4ef6\u3002\u5206\u6790\u8868\u660e\uff0cLoRA\u548cFT\u8868\u73b0\u51fa\u4e0d\u540c\u7684\u5b66\u4e60\u6a21\u5f0f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>LoRA\u663e\u793a\u51fa\u65b9\u5411\u548c\u5e45\u5ea6\u53d8\u5316\u4e4b\u95f4\u7684\u6bd4\u4f8b\u5173\u7cfb\uff0c\u7f3a\u4e4f\u7ec6\u5fae\u8c03\u6574\u80fd\u529b\u3002<\/li>\n\n\n\n<li>FT\u5c55\u793a\u4e86\u66f4\u4e3a\u591a\u6837\u7684\u5b66\u4e60\u6a21\u5f0f\uff0c\u5177\u6709\u8d1f\u659c\u7387\u8d8b\u52bf\uff0c\u8868\u660e\u5176\u5177\u5907\u8fdb\u884c\u7ec6\u5fae\u8c03\u6574\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u56db\u3001\u65b9\u6cd5\u8bba\uff1a\u6743\u91cd\u5206\u89e3\u4f4e\u79e9\u9002\u914d\uff08DoRA\uff09<\/h4>\n\n\n\n<p>DoRA\u5c06\u9884\u8bad\u7ec3\u6743\u91cd\u5206\u89e3\u4e3a\u5e45\u5ea6\uff08m\uff09\u548c\u65b9\u5411\uff08V\uff09\u7ec4\u4ef6\u3002\u65b9\u5411\u7ec4\u4ef6\u901a\u8fc7LoRA\u8fdb\u4e00\u6b65\u66f4\u65b0\uff0c\u5b9e\u73b0\u9ad8\u6548\u5fae\u8c03\u3002\u8fd9\u79cd\u5206\u89e3\u7b80\u5316\u4e86LoRA\u7684\u5b66\u4e60\u4efb\u52a1\uff0c\u4e13\u6ce8\u4e8e\u65b9\u5411\u66f4\u65b0\uff0c\u540c\u65f6\u901a\u8fc7\u6743\u91cd\u5f52\u4e00\u5316\u539f\u5219\u4fdd\u6301\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<p><strong>\u516c\u5f0f<\/strong>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"200\" height=\"47\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/07\/image-10.png\" alt=\"\" class=\"wp-image-4105\"\/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e94\u3001DoRA\u7684\u68af\u5ea6\u5206\u6790<\/h4>\n\n\n\n<p>DoRA\u7684\u68af\u5ea6\u5206\u6790\u663e\u793a\uff0c\u6743\u91cd\u5206\u89e3\u4f7f\u68af\u5ea6\u7684\u534f\u65b9\u5dee\u77e9\u9635\u66f4\u63a5\u8fd1\u4e8e\u5355\u4f4d\u77e9\u9635\uff0c\u6709\u5229\u4e8e\u4f18\u5316\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u66f4\u7a33\u5b9a\u7684\u5b66\u4e60\u6a21\u5f0f\uff0c\u6709\u6548\u589e\u5f3a\u4e86LoRA\u7684\u5b66\u4e60\u80fd\u529b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u516d\u3001\u51cf\u5c11\u8bad\u7ec3\u5f00\u9500<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u51cf\u5c11\u5185\u5b58\u6d88\u8017\uff0cDoRA\u5c06\u65b9\u5411\u66f4\u65b0\u7684\u8303\u6570\u89c6\u4e3a\u5e38\u6570\uff0c\u4ece\u68af\u5ea6\u56fe\u4e2d\u5206\u79bb\u51fa\u6765\u3002\u6b64\u8c03\u6574\u663e\u8457\u51cf\u5c11\u4e86\u5185\u5b58\u4f7f\u7528\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u5176\u3001\u5b9e\u9a8c<\/h4>\n\n\n\n<p>DoRA\u5728\u5404\u79cd\u4efb\u52a1\u548c\u6a21\u578b\u4e2d\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u663e\u793a\u51fa\u76f8\u5bf9\u4e8eLoRA\u7684\u4e00\u8d2f\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5e38\u8bc6\u63a8\u7406<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>DoRA\u5728LLaMA\u6a21\u578b\u4e0a\u663e\u8457\u4f18\u4e8eLoRA\u3002<\/li>\n\n\n\n<li>\u91c7\u7528\u51cf\u534a\u79e9\u7684DoRA\uff08DoRA\u2020\uff09\u4ecd\u4f18\u4e8eLoRA\uff0c\u7a81\u663e\u5176\u9ad8\u6548\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u56fe\u50cf\/\u89c6\u9891-\u6587\u672c\u7406\u89e3<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728VL-BART\u4e0a\uff0cDoRA\u5728\u56fe\u50cf\u548c\u89c6\u9891-\u6587\u672c\u4efb\u52a1\u4e2d\u5747\u4f18\u4e8eLoRA\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u89c6\u89c9\u6307\u4ee4\u5fae\u8c03<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728LLaVA-1.5-7B\u7684\u89c6\u89c9\u6307\u4ee4\u5fae\u8c03\u4efb\u52a1\u4e2d\uff0cDoRA\u8868\u73b0\u4f18\u4e8eLoRA\u548cFT\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u4e0e\u5176\u4ed6LoRA\u53d8\u79cd\u7684\u517c\u5bb9\u6027<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>DoRA\u4e0eVeRA\u517c\u5bb9\uff0c\u7ec4\u5408\u4f7f\u7528\uff08DVoRA\uff09\u8868\u73b0\u4f18\u4e8eVeRA\u548cLoRA\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u516b\u3001DoRA\u7684\u9c81\u68d2\u6027\u548c\u5fae\u8c03\u7c92\u5ea6<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>DoRA\u5728\u4e0d\u540c\u79e9\u548c\u8bad\u7ec3\u6837\u672c\u91cf\u4e0b\u5747\u8868\u73b0\u51fa\u8272\uff0c\u4e00\u76f4\u4f18\u4e8eLoRA\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7\u9009\u62e9\u6027\u5730\u4ec5\u66f4\u65b0\u7279\u5b9a\u6a21\u5757\u7684\u5e45\u5ea6\u7ec4\u4ef6\uff0cDoRA\u53ef\u4ee5\u5728\u51cf\u5c11\u53ef\u8bad\u7ec3\u53c2\u6570\u7684\u60c5\u51b5\u4e0b\u4fdd\u6301\u8f83\u9ad8\u7684\u51c6\u786e\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e5d\u3001\u66f4\u5e7f\u6cdb\u7684\u5f71\u54cd<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>QDoRA<\/strong>\uff1a\u5728QLoRA\u7684\u57fa\u7840\u4e0a\u589e\u5f3aLoRA\u7684\u51c6\u786e\u6027\uff0c\u663e\u793a\u51fa\u5728Orca-Math\u4e0a\u7684\u826f\u597d\u7ed3\u679c\u3002<\/li>\n\n\n\n<li><strong>\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210<\/strong>\uff1aDoRA\u5728\u7ec6\u5316\u7a33\u5b9a\u6269\u6563\u6a21\u578b\u7684\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u4e2d\u663e\u8457\u4f18\u4e8eLoRA\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u5341\u3001\u7ed3\u8bba<\/h4>\n\n\n\n<p>DoRA\u7f29\u5c0f\u4e86LoRA\u548cFT\u4e4b\u95f4\u7684\u5dee\u8ddd\uff0c\u63d0\u4f9b\u4e86\u4e00\u79cd\u53c2\u6570\u9ad8\u6548\u7684\u5fae\u8c03\u65b9\u6cd5\uff0c\u5176\u5b66\u4e60\u80fd\u529b\u63a5\u8fd1\u4e8eFT\u3002DoRA\u5728\u591a\u4e2a\u4efb\u52a1\u548c\u6a21\u578b\u67b6\u6784\u4e2d\u8868\u73b0\u51fa\u6301\u7eed\u6539\u8fdb\uff0c\u4e14\u65e0\u989d\u5916\u7684\u63a8\u7406\u5f00\u9500\u3002\u672a\u6765\u5de5\u4f5c\u5c06\u63a2\u7d22DoRA\u5728\u97f3\u9891\u7b49\u5176\u4ed6\u9886\u57df\u7684\u9002\u7528\u6027\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>DoRA on GitHub: <a href=\"https:\/\/github.com\/NVlabs\/DoRA\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/NVlabs\/DoRA<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587DoRA: Weight-Decomposed Low-Rank Adaptation\uff08\u6743\u91cd\u5206\u89e3\u4f4e\u79e9\u9002\u914d\uff09 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8,36],"tags":[39,84],"class_list":["post-4104","post","type-post","status-publish","format-standard","hentry","category-tech","category-36","tag-ai","tag-nvidia"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a> <a href=\"https:\/\/nullthought.net\/?cat=36\" rel=\"category\">\u79d1\u5b66<\/a>","rttpg_excerpt":"\u8bba\u6587DoRA: Weight-Decomposed Low-Rank Adaptation\uff08\u6743\u91cd\u5206\u89e3\u4f4e\u79e9\u9002\u914d\uff09&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4104","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4104"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4104\/revisions"}],"predecessor-version":[{"id":4108,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4104\/revisions\/4108"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4104"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4104"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4104"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}