{"id":83911,"date":"2026-07-26T06:17:00","date_gmt":"2026-07-25T22:17:00","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/83911.html"},"modified":"2026-07-26T06:17:00","modified_gmt":"2026-07-25T22:17:00","slug":"peft-2-0%e8%bf%9b%e9%98%b6%ef%bc%9aubuntu%e6%9c%8d%e5%8a%a1%e5%99%a8%e4%b8%8a%e7%9a%84%e9%ab%98%e6%95%88%e5%be%ae%e8%b0%83%e7%ad%96%e7%95%a5%e4%b8%8e%e4%bc%98%e5%8c%96","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/83911.html","title":{"rendered":"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316"},"content":{"rendered":"<p>\u4f5c\u8005&#xff1a;\u5434\u4e1a\u4eae<br \/>\n\u535a\u5ba2&#xff1a;wuyeliang.blog.csdn.net<\/p>\n<h3>1 PEFT\u6280\u672f\u6838\u5fc3\u539f\u7406<\/h3>\n<p>\u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;Parameter-Efficient Fine-Tuning, PEFT&#xff09;\u662f\u4e00\u79cd\u9488\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b&#xff08;\u5c24\u5176\u662f\u5927\u8bed\u8a00\u6a21\u578b&#xff09;\u7684\u5fae\u8c03\u7b56\u7565&#xff0c;\u5176\u6838\u5fc3\u601d\u60f3\u662f\u907f\u514d\u5bf9\u6a21\u578b\u7684\u5168\u90e8\u53c2\u6570\u8fdb\u884c\u66f4\u65b0&#xff0c;\u800c\u662f\u4ec5\u8c03\u6574\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u6216\u5f15\u5165\u5c11\u91cf\u989d\u5916\u53c2\u6570&#xff0c;\u4ece\u800c\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u548c\u5b58\u50a8\u6210\u672c&#xff0c;\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<h4>1.1 PEFT\u7684\u5b58\u5728\u610f\u4e49<\/h4>\n<p>\u4f20\u7edf\u5168\u53c2\u6570\u5fae\u8c03\u9700\u8981\u66f4\u65b0\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570&#xff0c;\u8fd9\u5bf9\u4e8e\u6570\u5341\u4ebf\u751a\u81f3\u5343\u4ebf\u53c2\u6570\u7684\u5927\u6a21\u578b\u6765\u8bf4&#xff0c;\u9762\u4e34\u4e09\u5927\u6311\u6218&#xff1a;<\/p>\n<ul>\n<li>\u8ba1\u7b97\u6210\u672c\u9ad8\u6602&#xff1a;\u5168\u53c2\u6570\u5fae\u8c03\u4e00\u4e2a12B\u53c2\u6570\u7684\u6a21\u578b\u9700\u8981\u7ea680GB\u663e\u5b58&#xff0c;\u800c\u4f7f\u7528LoRA\u7b49PEFT\u6280\u672f\u4ec5\u970018GB\u5de6\u53f3<\/li>\n<li>\u707e\u96be\u6027\u9057\u5fd8&#xff1a;\u5168\u53c2\u6570\u5fae\u8c03\u53ef\u80fd\u5bfc\u81f4\u9884\u8bad\u7ec3\u77e5\u8bc6\u4e22\u5931&#xff0c;PEFT\u901a\u8fc7\u51bb\u7ed3\u4e3b\u4f53\u53c2\u6570\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898<\/li>\n<li>\u5b58\u50a8\u5f00\u9500\u5927&#xff1a;\u6bcf\u4e2a\u4efb\u52a1\u90fd\u9700\u8981\u4fdd\u5b58\u5b8c\u6574\u7684\u6a21\u578b\u526f\u672c&#xff0c;\u800cPEFT\u53ea\u9700\u4fdd\u5b58\u5c11\u91cf\u9002\u914d\u53c2\u6570<\/li>\n<\/ul>\n<h4>1.2 \u4e3b\u6d41PEFT\u65b9\u6cd5\u5206\u7c7b<\/h4>\n<p>PEFT\u65b9\u6cd5\u4e3b\u8981\u5206\u4e3a\u4e09\u5927\u7c7b\u522b&#xff1a;<\/p>\n<h5>1.2.1 \u589e\u52a0\u5f0f\u65b9\u6cd5&#xff08;Additive Methods&#xff09;<\/h5>\n<p>\u8fd9\u7c7b\u65b9\u6cd5\u901a\u8fc7\u589e\u52a0\u989d\u5916\u7684\u53c2\u6570\u6216\u5c42\u6765\u6269\u5c55\u73b0\u6709\u9884\u8bad\u7ec3\u6a21\u578b&#xff0c;\u4ec5\u8bad\u7ec3\u65b0\u589e\u52a0\u7684\u53c2\u6570&#xff1a;<\/p>\n<ul>\n<li>Adapter Tuning&#xff1a;\u5728Transformer\u5c42\u4e2d\u63d2\u5165\u5c0f\u578b\u795e\u7ecf\u7f51\u7edc\u6a21\u5757<\/li>\n<li>\u8f6f\u63d0\u793a\u65b9\u6cd5&#xff1a;\u5982Prefix Tuning\u3001Prompt Tuning\u3001P-Tuning\u7b49<\/li>\n<\/ul>\n<h5>1.2.2 \u9009\u62e9\u5f0f\u65b9\u6cd5&#xff08;Selective Methods&#xff09;<\/h5>\n<p>\u9009\u62e9\u6a21\u578b\u7684\u73b0\u6709\u53c2\u6570\u5b50\u96c6\u8fdb\u884c\u5fae\u8c03&#xff1a;<\/p>\n<ul>\n<li>BitFit&#xff1a;\u4ec5\u5fae\u8c03\u6a21\u578b\u7684\u504f\u7f6e&#xff08;bias&#xff09;\u53c2\u6570&#xff0c;\u53c2\u6570\u91cf\u5360\u6bd4\u4e0d\u52300.1%<\/li>\n<\/ul>\n<h5>1.2.3 \u91cd\u65b0\u53c2\u6570\u5316\u65b9\u6cd5&#xff08;Reparameterization Methods&#xff09;<\/h5>\n<p>\u5229\u7528\u4f4e\u79e9\u8868\u793a\u6700\u5c0f\u5316\u53ef\u8bad\u7ec3\u53c2\u6570\u6570\u91cf&#xff1a;<\/p>\n<ul>\n<li>LoRA&#xff1a;\u901a\u8fc7\u4f4e\u79e9\u5206\u89e3\u4f18\u5316\u6743\u91cd\u77e9\u9635\u7684\u589e\u91cf\u66f4\u65b0<\/li>\n<li>AdaLoRA&#xff1a;LoRA\u7684\u5347\u7ea7\u7248&#xff0c;\u81ea\u9002\u5e94\u8c03\u6574\u4e0d\u540c\u6a21\u5757\u7684\u79e9<\/li>\n<\/ul>\n<p>\u8868&#xff1a;\u4e3b\u8981PEFT\u65b9\u6cd5\u5bf9\u6bd4\u5206\u6790<\/p>\n<table>\n<tr>\u65b9\u6cd5\u53c2\u6570\u6548\u7387\u63a8\u7406\u5ef6\u8fdf\u6027\u80fd\u8868\u73b0\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>Adapter<\/td>\n<td>\u9ad8<\/td>\n<td>\u8f7b\u5fae\u589e\u52a0<\/td>\n<td>\u63a5\u8fd1\u5168\u53c2\u6570\u5fae\u8c03<\/td>\n<td>\u591a\u4efb\u52a1\u3001\u8de8\u8bed\u8a00\u8fc1\u79fb<\/td>\n<\/tr>\n<tr>\n<td>LoRA<\/td>\n<td>\u6781\u9ad8<\/td>\n<td>\u65e0\u589e\u52a0<\/td>\n<td>\u63a5\u8fd1\u5168\u53c2\u6570\u5fae\u8c03<\/td>\n<td>\u5927\u6a21\u578b\u9002\u914d\u3001\u4e2a\u6027\u5316<\/td>\n<\/tr>\n<tr>\n<td>Prefix Tuning<\/td>\n<td>\u6781\u9ad8<\/td>\n<td>\u65e0\u589e\u52a0<\/td>\n<td>\u7a0d\u900a\u4e8e\u5168\u53c2\u6570\u5fae\u8c03<\/td>\n<td>\u751f\u6210\u4efb\u52a1\u3001\u4f4e\u8d44\u6e90\u573a\u666f<\/td>\n<\/tr>\n<tr>\n<td>Prompt Tuning<\/td>\n<td>\u6781\u9ad8<\/td>\n<td>\u65e0\u589e\u52a0<\/td>\n<td>\u7a0d\u900a\u4e8e\u5168\u53c2\u6570\u5fae\u8c03<\/td>\n<td>\u5c0f\u6570\u636e\u96c6\u3001API\u6a21\u578b<\/td>\n<\/tr>\n<tr>\n<td>Sparse Fine-Tuning<\/td>\n<td>\u9ad8<\/td>\n<td>\u65e0\u589e\u52a0<\/td>\n<td>\u4f9d\u8d56\u7b56\u7565\u8d28\u91cf<\/td>\n<td>\u6a21\u578b\u538b\u7f29\u3001\u8d44\u6e90\u53d7\u9650<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>2 Ubuntu 22.04\u73af\u5883\u914d\u7f6e<\/h3>\n<p>\u5728\u5f00\u59cbPEFT\u5b9e\u8df5\u524d&#xff0c;\u9700\u8981\u5728Ubuntu 22.04\u7cfb\u7edf\u4e0a\u914d\u7f6e\u5408\u9002\u7684\u5f00\u53d1\u73af\u5883\u3002<\/p>\n<h4>2.1 \u7cfb\u7edf\u57fa\u7840\u8981\u6c42<\/h4>\n<ul>\n<li>\u64cd\u4f5c\u7cfb\u7edf&#xff1a;Ubuntu 22.04 LTS<\/li>\n<li>Python\u7248\u672c&#xff1a;Python 3.8\u53ca\u4ee5\u4e0a&#xff08;\u63a8\u83503.10&#xff09;<\/li>\n<li>CUDA\u652f\u6301&#xff1a;CUDA 12.1<\/li>\n<li>PyTorch&#xff1a;PyTorch 2.3.0\u53ca\u4ee5\u4e0a<\/li>\n<\/ul>\n<h4>2.2 \u6df1\u5ea6\u5b66\u4e60\u73af\u5883\u914d\u7f6e<\/h4>\n<p>\u4ee5\u4e0b\u662f\u5b8c\u6574\u7684\u73af\u5883\u914d\u7f6e\u6b65\u9aa4&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u66f4\u65b0\u7cfb\u7edf\u5305\u7ba1\u7406\u5668<\/span><br \/>\n<span class=\"token function\">sudo<\/span> <span class=\"token function\">apt<\/span> update<br \/>\n<span class=\"token function\">sudo<\/span> <span class=\"token function\">apt<\/span> upgrade -y<\/p>\n<p><span class=\"token comment\"># \u5b89\u88c5Python\u5f00\u53d1\u5de5\u5177<\/span><br \/>\n<span class=\"token function\">sudo<\/span> <span class=\"token function\">apt<\/span> <span class=\"token function\">install<\/span> python3-pip python3-venv python3-dev -y<\/p>\n<p><span class=\"token comment\"># \u5b89\u88c5CUDA\u5de5\u5177\u5305&#xff08;\u5982\u65e0NVIDIA\u9a71\u52a8\u9700\u5148\u5b89\u88c5&#xff09;<\/span><br \/>\n<span class=\"token function\">sudo<\/span> <span class=\"token function\">apt<\/span> <span class=\"token function\">install<\/span> nvidia-cuda-toolkit -y<\/p>\n<p><span class=\"token comment\"># \u521b\u5efaPython\u865a\u62df\u73af\u5883<\/span><br \/>\npython3 -m venv peft_env<br \/>\n<span class=\"token builtin class-name\">source<\/span> peft_env\/bin\/activate<\/p>\n<p><span class=\"token comment\"># \u914d\u7f6ePyPI\u955c\u50cf\u6e90\u52a0\u901f\u4e0b\u8f7d<\/span><br \/>\npip config <span class=\"token builtin class-name\">set<\/span> global.index-url https:\/\/pypi.tuna.tsinghua.edu.cn\/simple<\/p>\n<p><span class=\"token comment\"># \u5b89\u88c5\u57fa\u7840\u6df1\u5ea6\u5b66\u4e60\u5e93<\/span><br \/>\npip <span class=\"token function\">install<\/span> &#8211;upgrade pip<br \/>\npip <span class=\"token function\">install<\/span> <span class=\"token assign-left variable\">torch<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">2.3<\/span>.0&#043;cu121 <span class=\"token assign-left variable\">torchvision<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">0.18<\/span>.0&#043;cu121 &#8211;extra-index-url https:\/\/download.pytorch.org\/whl\/cu121<br \/>\npip <span class=\"token function\">install<\/span> <span class=\"token assign-left variable\">transformers<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">4.43<\/span>.2 <span class=\"token assign-left variable\">datasets<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">2.20<\/span>.0 <span class=\"token assign-left variable\">accelerate<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">0.32<\/span>.1<\/p>\n<p><span class=\"token comment\"># \u5b89\u88c5PEFT\u5e93<\/span><br \/>\npip <span class=\"token function\">install<\/span> <span class=\"token assign-left variable\">peft<\/span><span class=\"token operator\">&#061;&#061;<\/span><span class=\"token number\">0.11<\/span>.1<\/p>\n<p><span class=\"token comment\"># \u5b89\u88c5\u8bad\u7ec3\u6240\u9700\u5176\u4ed6\u4f9d\u8d56<\/span><br \/>\npip <span class=\"token function\">install<\/span> tensorboard scikit-learn jupyter<\/p>\n<h4>2.3 \u73af\u5883\u9a8c\u8bc1<\/h4>\n<p>\u521b\u5efa\u6d4b\u8bd5\u811a\u672c\u9a8c\u8bc1\u73af\u5883\u662f\u5426\u6b63\u786e\u914d\u7f6e&#xff1a;<\/p>\n<p><span class=\"token comment\"># test_environment.py<\/span><br \/>\n<span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">import<\/span> transformers<br \/>\n<span class=\"token keyword\">from<\/span> peft <span class=\"token keyword\">import<\/span> LoraConfig<span class=\"token punctuation\">,<\/span> get_peft_model<br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;PyTorch\u7248\u672c: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>torch<span class=\"token punctuation\">.<\/span>__version__<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;CUDA\u53ef\u7528: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>torch<span class=\"token punctuation\">.<\/span>cuda<span class=\"token punctuation\">.<\/span>is_available<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;CUDA\u7248\u672c: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>torch<span class=\"token punctuation\">.<\/span>version<span class=\"token punctuation\">.<\/span>cuda<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;GPU\u6570\u91cf: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>torch<span class=\"token punctuation\">.<\/span>cuda<span class=\"token punctuation\">.<\/span>device_count<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">if<\/span> torch<span class=\"token punctuation\">.<\/span>cuda<span class=\"token punctuation\">.<\/span>is_available<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;\u5f53\u524dGPU: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>torch<span class=\"token punctuation\">.<\/span>cuda<span class=\"token punctuation\">.<\/span>get_device_name<span class=\"token punctuation\">(<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;Transformers\u7248\u672c: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>transformers<span class=\"token punctuation\">.<\/span>__version__<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u73af\u5883\u914d\u7f6e\u6210\u529f&#xff01;&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h3>3 LoRA\u5fae\u8c03\u5b9e\u8df5\u8be6\u89e3<\/h3>\n<p>LoRA&#xff08;Low-Rank Adaptation&#xff09;\u662f\u76ee\u524d\u6700\u6d41\u884c\u7684PEFT\u65b9\u6cd5\u4e4b\u4e00&#xff0c;\u4e0b\u9762\u4ee5\u5177\u4f53\u6848\u4f8b\u5c55\u793a\u5176\u5728Ubuntu 22.04\u4e0a\u7684\u5b9e\u73b0\u3002<\/p>\n<h4>3.1 LoRA\u6838\u5fc3\u539f\u7406<\/h4>\n<p>LoRA\u57fa\u4e8e\u4e00\u4e2a\u5173\u952e\u5047\u8bbe&#xff1a;\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u4efb\u52a1\u9002\u914d\u8fc7\u7a0b\u4e2d\u6743\u91cd\u7684\u6539\u53d8\u91cf\u662f\u4f4e\u79e9\u7684\u3002\u5176\u6570\u5b66\u8868\u8fbe\u4e3a&#xff1a;<\/p>\n<p>\u5bf9\u4e8e\u539f\u59cb\u6743\u91cd\u77e9\u9635 ( W \\\\in \\\\mathbb{R}^{d \\\\times k} )&#xff0c;\u6743\u91cd\u66f4\u65b0\u53ef\u8868\u793a\u4e3a\u4f4e\u79e9\u5f62\u5f0f&#xff1a;<br \/>\n[ \\\\Delta W &#061; A \\\\cdot B ]<br \/>\n\u5176\u4e2d ( A \\\\in \\\\mathbb{R}^{d \\\\times r} ), ( B \\\\in \\\\mathbb{R}^{r \\\\times k} ), \u4e14 ( r \\\\ll \\\\min(d, k) ) \u3002<\/p>\n<p>\u8bad\u7ec3\u65f6\u51bb\u7ed3\u539f\u59cb\u6743\u91cd ( W )&#xff0c;\u53ea\u66f4\u65b0 ( A ) \u548c ( B )\u3002\u63a8\u7406\u65f6\u53ef\u5c06 ( \\\\Delta W ) \u5408\u5e76\u5230 ( W ) \u4e2d&#xff0c;\u4e0d\u4ea7\u751f\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n<h4>3.2 \u5173\u952e\u53c2\u6570\u914d\u7f6e<\/h4>\n<p>LoRA\u7684\u6027\u80fd\u9ad8\u5ea6\u4f9d\u8d56\u4e09\u4e2a\u6838\u5fc3\u53c2\u6570\u7684\u5408\u7406\u914d\u7f6e&#xff1a;<\/p>\n<h5>3.2.1 \u5b66\u4e60\u7387&#xff08;Learning Rate&#xff09;<\/h5>\n<p>PEFT\u8bad\u7ec3\u901a\u5e38\u9700\u8981\u6bd4\u5168\u53c2\u6570\u5fae\u8c03\u66f4\u5927\u7684\u5b66\u4e60\u7387\u3002\u57fa\u7840\u5b66\u4e60\u7387\u53ef\u53c2\u8003\u516c\u5f0f&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u57fa\u7840\u5b66\u4e60\u7387\u8ba1\u7b97\u516c\u5f0f&#xff08;\u9002\u7528\u4e8eLlama\/OPT\u7b49\u4e3b\u6d41\u6a21\u578b&#xff09;<\/span><br \/>\nlearning_rate <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">2e<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">4<\/span> <span class=\"token operator\">\/<\/span> <span class=\"token punctuation\">(<\/span>model_dimension <span class=\"token operator\">\/<\/span> <span class=\"token number\">768<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u4f8b\u5982&#xff0c;768\u7ef4\u6a21\u578b&#xff08;\u5982BERT-base&#xff09;\u4f7f\u75282e-4&#xff0c;3072\u7ef4\u6a21\u578b&#xff08;\u5982Llama-3.2-3B&#xff09;\u4f7f\u75285e-5\u3002<\/p>\n<h5>3.2.2 \u79e9&#xff08;Rank&#xff09;<\/h5>\n<p>\u79e9\u63a7\u5236\u4f4e\u79e9\u77e9\u9635\u7684\u7ef4\u5ea6&#xff0c;\u5f71\u54cd\u53ef\u8bad\u7ec3\u53c2\u6570\u6570\u91cf\u548c\u6a21\u578b\u8868\u8fbe\u80fd\u529b&#xff1a;<\/p>\n<ul>\n<li>\u5c0f\u6a21\u578b&#xff08;&lt;1B\u53c2\u6570&#xff09;&#xff1a;\u63a8\u8350\u79e9&#061;16<\/li>\n<li>\u4e2d\u6a21\u578b&#xff08;1B-10B\u53c2\u6570&#xff09;&#xff1a;\u63a8\u8350\u79e9&#061;32<\/li>\n<li>\u5927\u6a21\u578b&#xff08;&gt;10B\u53c2\u6570&#xff09;&#xff1a;\u63a8\u8350\u79e9&#061;64<\/li>\n<\/ul>\n<h5>3.2.3 Alpha\u53c2\u6570<\/h5>\n<p>Alpha\u662f\u7f29\u653e\u56e0\u5b50&#xff0c;\u63a7\u5236\u4f4e\u79e9\u77e9\u9635\u5bf9\u8f93\u51fa\u7684\u5f71\u54cd\u7a0b\u5ea6\u3002\u5b9e\u9a8c\u8868\u660e\u6700\u4f73\u914d\u7f6e\u4e3a&#xff1a;<\/p>\n<p>alpha <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">2<\/span> <span class=\"token operator\">*<\/span> rank  <span class=\"token comment\"># \u5f53rank&#061;32\u65f6&#xff0c;alpha&#061;64<\/span><\/p>\n<p>\u8fd9\u4e00\u6bd4\u4f8b\u5728LoRA\u3001LoHa\u3001LoKr\u7b49\u591a\u79cdPEFT\u65b9\u6cd5\u4e2d\u5f97\u5230\u9a8c\u8bc1\u3002<\/p>\n<h4>3.3 \u5b8c\u6574\u5b9e\u8df5\u6848\u4f8b<\/h4>\n<p>\u4ee5\u4e0b\u662f\u5728Ubuntu 22.04\u4e0a\u4f7f\u7528LoRA\u5fae\u8c03\u4ee3\u7801\u751f\u6210\u6a21\u578b\u7684\u5b8c\u6574\u793a\u4f8b&#xff1a;<\/p>\n<p><span class=\"token comment\"># lora_finetuning.py<\/span><br \/>\n<span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> AutoModelForCausalLM<span class=\"token punctuation\">,<\/span> AutoTokenizer<span class=\"token punctuation\">,<\/span> TrainingArguments<span class=\"token punctuation\">,<\/span> Trainer<br \/>\n<span class=\"token keyword\">from<\/span> peft <span class=\"token keyword\">import<\/span> LoraConfig<span class=\"token punctuation\">,<\/span> get_peft_model<span class=\"token punctuation\">,<\/span> TaskType<br \/>\n<span class=\"token keyword\">from<\/span> datasets <span class=\"token keyword\">import<\/span> load_dataset<br \/>\n<span class=\"token keyword\">import<\/span> os<\/p>\n<p><span class=\"token comment\"># 1. \u52a0\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b\u548c\u5206\u8bcd\u5668<\/span><br \/>\nmodel_name <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#034;deepseek-ai\/DeepSeek-Coder-V2-Lite-Instruct&#034;<\/span><br \/>\ntokenizer <span class=\"token operator\">&#061;<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>model_name<span class=\"token punctuation\">,<\/span> use_fast<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">,<\/span> trust_remote_code<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><br \/>\ntokenizer<span class=\"token punctuation\">.<\/span>padding_side <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#039;right&#039;<\/span>  <span class=\"token comment\"># \u8bbe\u7f6e\u53f3\u586b\u5145\u6a21\u5f0f<\/span><\/p>\n<p>model <span class=\"token operator\">&#061;<\/span> AutoModelForCausalLM<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span><br \/>\n    model_name<span class=\"token punctuation\">,<\/span><br \/>\n    trust_remote_code<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    torch_dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">,<\/span><br \/>\n    device_map<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;auto&#034;<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># 2. \u914d\u7f6eLoRA\u53c2\u6570<\/span><br \/>\nlora_config <span class=\"token operator\">&#061;<\/span> LoraConfig<span class=\"token punctuation\">(<\/span><br \/>\n    task_type<span class=\"token operator\">&#061;<\/span>TaskType<span class=\"token punctuation\">.<\/span>CAUSAL_LM<span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token comment\"># \u76ee\u6807\u5fae\u8c03\u5c42&#xff1a;\u6ce8\u610f\u529b\u673a\u5236\u4e0eFFN\u6a21\u5757<\/span><br \/>\n    target_modules<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;q_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;kv_a_proj_with_mqa&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;kv_b_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;o_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;gate_proj&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;up_proj&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;down_proj&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    inference_mode<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u8bad\u7ec3\u6a21\u5f0f\u8bbe\u7f6e<\/span><br \/>\n    r<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u4f4e\u79e9\u77e9\u9635\u7ef4\u5ea6<\/span><br \/>\n    lora_alpha<span class=\"token operator\">&#061;<\/span><span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u7f29\u653e\u56e0\u5b50<\/span><br \/>\n    lora_dropout<span class=\"token operator\">&#061;<\/span><span class=\"token number\">0.1<\/span>  <span class=\"token comment\"># \u6b63\u5219\u5316 dropout\u6bd4\u4f8b<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># 3. \u5c06\u6a21\u578b\u5305\u88c5\u4e3aPEFT\u6a21\u578b<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> get_peft_model<span class=\"token punctuation\">(<\/span>model<span class=\"token punctuation\">,<\/span> lora_config<span class=\"token punctuation\">)<\/span><br \/>\nmodel<span class=\"token punctuation\">.<\/span>print_trainable_parameters<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># \u6253\u5370\u53ef\u8bad\u7ec3\u53c2\u6570\u6bd4\u4f8b<\/span><\/p>\n<p><span class=\"token comment\"># 4. \u52a0\u8f7d\u5e76\u9884\u5904\u7406\u6570\u636e\u96c6<\/span><br \/>\n<span class=\"token keyword\">def<\/span> <span class=\"token function\">preprocess_function<\/span><span class=\"token punctuation\">(<\/span>examples<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token comment\"># \u6784\u5efa\u5bf9\u8bdd\u683c\u5f0f<\/span><br \/>\n    instructions <span class=\"token operator\">&#061;<\/span> examples<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;instruction&#034;<\/span><span class=\"token punctuation\">]<\/span><br \/>\n    inputs <span class=\"token operator\">&#061;<\/span> examples<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;input&#034;<\/span><span class=\"token punctuation\">]<\/span><br \/>\n    outputs <span class=\"token operator\">&#061;<\/span> examples<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;output&#034;<\/span><span class=\"token punctuation\">]<\/span><\/p>\n<p>    texts <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">]<\/span><br \/>\n    <span class=\"token keyword\">for<\/span> instr<span class=\"token punctuation\">,<\/span> inp<span class=\"token punctuation\">,<\/span> out <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">zip<\/span><span class=\"token punctuation\">(<\/span>instructions<span class=\"token punctuation\">,<\/span> inputs<span class=\"token punctuation\">,<\/span> outputs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> inp<span class=\"token punctuation\">:<\/span><br \/>\n            text <span class=\"token operator\">&#061;<\/span> <span class=\"token string-interpolation\"><span class=\"token string\">f&#034;Instruction: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>instr<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">\\\\nInput: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>inp<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">\\\\nOutput: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>out<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><br \/>\n        <span class=\"token keyword\">else<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            text <span class=\"token operator\">&#061;<\/span> <span class=\"token string-interpolation\"><span class=\"token string\">f&#034;Instruction: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>instr<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">\\\\nOutput: <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>out<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#034;<\/span><\/span><br \/>\n        texts<span class=\"token punctuation\">.<\/span>append<span class=\"token punctuation\">(<\/span>text<span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token comment\"># \u4ee4\u724c\u5316<\/span><br \/>\n    tokenized <span class=\"token operator\">&#061;<\/span> tokenizer<span class=\"token punctuation\">(<\/span>texts<span class=\"token punctuation\">,<\/span> truncation<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span> max_length<span class=\"token operator\">&#061;<\/span><span class=\"token number\">384<\/span><span class=\"token punctuation\">,<\/span> padding<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;max_length&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    tokenized<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;labels&#034;<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> tokenized<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;input_ids&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>copy<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> tokenized<\/p>\n<p><span class=\"token comment\"># \u52a0\u8f7d\u793a\u4f8b\u6570\u636e\u96c6&#xff08;\u5b9e\u9645\u4f7f\u7528\u65f6\u53ef\u66ff\u6362\u4e3a\u81ea\u5b9a\u4e49\u6570\u636e&#xff09;<\/span><br \/>\ndataset <span class=\"token operator\">&#061;<\/span> load_dataset<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;json&#034;<\/span><span class=\"token punctuation\">,<\/span> data_files<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/dataset\/train.json&#034;<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># \u5047\u8bbe\u6570\u636e\u6587\u4ef6\u8def\u5f84<\/span><br \/>\ntokenized_dataset <span class=\"token operator\">&#061;<\/span> dataset<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">map<\/span><span class=\"token punctuation\">(<\/span>preprocess_function<span class=\"token punctuation\">,<\/span> batched<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># 5. \u914d\u7f6e\u8bad\u7ec3\u53c2\u6570<\/span><br \/>\ntraining_args <span class=\"token operator\">&#061;<\/span> TrainingArguments<span class=\"token punctuation\">(<\/span><br \/>\n    output_dir<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/output\/deepseek_coder_v2&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    per_device_train_batch_size<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    gradient_accumulation_steps<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    logging_steps<span class=\"token operator\">&#061;<\/span><span class=\"token number\">10<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    num_train_epochs<span class=\"token operator\">&#061;<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    save_steps<span class=\"token operator\">&#061;<\/span><span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1e<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">5<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    save_on_each_node<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    gradient_checkpointing<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\u8282\u7701\u663e\u5b58<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># 6. \u521b\u5efaTrainer\u5e76\u5f00\u59cb\u8bad\u7ec3<\/span><br \/>\ntrainer <span class=\"token operator\">&#061;<\/span> Trainer<span class=\"token punctuation\">(<\/span><br \/>\n    model<span class=\"token operator\">&#061;<\/span>model<span class=\"token punctuation\">,<\/span><br \/>\n    args<span class=\"token operator\">&#061;<\/span>training_args<span class=\"token punctuation\">,<\/span><br \/>\n    train_dataset<span class=\"token operator\">&#061;<\/span>tokenized_dataset<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;train&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    data_collator<span class=\"token operator\">&#061;<\/span><span class=\"token keyword\">lambda<\/span> data<span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n        <span class=\"token string\">&#039;input_ids&#039;<\/span><span class=\"token punctuation\">:<\/span> torch<span class=\"token punctuation\">.<\/span>stack<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span>torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>f<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;input_ids&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token keyword\">for<\/span> f <span class=\"token keyword\">in<\/span> data<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n        <span class=\"token string\">&#039;attention_mask&#039;<\/span><span class=\"token punctuation\">:<\/span> torch<span class=\"token punctuation\">.<\/span>stack<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span>torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>f<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;attention_mask&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token keyword\">for<\/span> f <span class=\"token keyword\">in<\/span> data<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n        <span class=\"token string\">&#039;labels&#039;<\/span><span class=\"token punctuation\">:<\/span> torch<span class=\"token punctuation\">.<\/span>stack<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span>torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>f<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;labels&#039;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token keyword\">for<\/span> f <span class=\"token keyword\">in<\/span> data<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token punctuation\">}<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>trainer<span class=\"token punctuation\">.<\/span>train<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># 7. \u4fdd\u5b58\u5fae\u8c03\u540e\u7684\u6a21\u578b<\/span><br \/>\nmodel<span class=\"token punctuation\">.<\/span>save_pretrained<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;.\/output\/deepseek_coder_v2_lora&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h4>3.4 \u6a21\u578b\u63a8\u7406\u4e0e\u5e94\u7528<\/h4>\n<p>\u8bad\u7ec3\u5b8c\u6210\u540e&#xff0c;\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u52a0\u8f7d\u5fae\u8c03\u540e\u7684\u6a21\u578b\u8fdb\u884c\u63a8\u7406&#xff1a;<\/p>\n<p><span class=\"token comment\"># inference.py<\/span><br \/>\n<span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> AutoModelForCausalLM<span class=\"token punctuation\">,<\/span> AutoTokenizer<br \/>\n<span class=\"token keyword\">from<\/span> peft <span class=\"token keyword\">import<\/span> PeftModel<br \/>\n<span class=\"token keyword\">import<\/span> torch<\/p>\n<p><span class=\"token comment\"># \u8def\u5f84\u914d\u7f6e<\/span><br \/>\nmodel_path <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#039;deepseek-ai\/DeepSeek-Coder-V2-Lite-Instruct&#039;<\/span><br \/>\nlora_path <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#039;.\/output\/deepseek_coder_v2_lora&#039;<\/span><\/p>\n<p><span class=\"token comment\"># \u52a0\u8f7d\u57fa\u7840\u6a21\u578b\u548c\u5206\u8bcd\u5668<\/span><br \/>\ntokenizer <span class=\"token operator\">&#061;<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>model_path<span class=\"token punctuation\">,<\/span> trust_remote_code<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> AutoModelForCausalLM<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span><br \/>\n    model_path<span class=\"token punctuation\">,<\/span><br \/>\n    device_map<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;auto&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    torch_dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span>bfloat16<span class=\"token punctuation\">,<\/span><br \/>\n    trust_remote_code<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># \u5408\u5e76LoRA\u6743\u91cd<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> PeftModel<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>model<span class=\"token punctuation\">,<\/span> model_id<span class=\"token operator\">&#061;<\/span>lora_path<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># \u6784\u5efa\u5bf9\u8bdd\u5386\u53f2<\/span><br \/>\nmessages <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><br \/>\n    <span class=\"token punctuation\">{<\/span><span class=\"token string\">&#039;role&#039;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#039;system&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;content&#039;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;\u4f60\u662f\u4e00\u4e2a\u6709\u7528\u7684\u7f16\u7a0b\u52a9\u624b&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token punctuation\">{<\/span><span class=\"token string\">&#039;role&#039;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#039;user&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;content&#039;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;\u7528Python\u5199\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u51fd\u6570&#034;<\/span><span class=\"token punctuation\">}<\/span><br \/>\n<span class=\"token punctuation\">]<\/span><\/p>\n<p><span class=\"token comment\"># \u751f\u6210\u56de\u590d<\/span><br \/>\ninputs <span class=\"token operator\">&#061;<\/span> tokenizer<span class=\"token punctuation\">.<\/span>apply_chat_template<span class=\"token punctuation\">(<\/span><br \/>\n    messages<span class=\"token punctuation\">,<\/span><br \/>\n    add_generation_prompt<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    return_tensors<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;pt&#034;<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>to<span class=\"token punctuation\">(<\/span>model<span class=\"token punctuation\">.<\/span>device<span class=\"token punctuation\">)<\/span><\/p>\n<p>outputs <span class=\"token operator\">&#061;<\/span> model<span class=\"token punctuation\">.<\/span>generate<span class=\"token punctuation\">(<\/span><br \/>\n    inputs<span class=\"token punctuation\">,<\/span><br \/>\n    max_new_tokens<span class=\"token operator\">&#061;<\/span><span class=\"token number\">512<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    do_sample<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    top_k<span class=\"token operator\">&#061;<\/span><span class=\"token number\">50<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    top_p<span class=\"token operator\">&#061;<\/span><span class=\"token number\">0.95<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    num_return_sequences<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    eos_token_id<span class=\"token operator\">&#061;<\/span>tokenizer<span class=\"token punctuation\">.<\/span>eos_token_id<br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>tokenizer<span class=\"token punctuation\">.<\/span>decode<span class=\"token punctuation\">(<\/span>outputs<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span><span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>inputs<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> skip_special_tokens<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<h3>4 \u4e0d\u540c\u573a\u666f\u4e0b\u7684PEFT\u914d\u7f6e\u7b56\u7565<\/h3>\n<p>\u6839\u636e\u5177\u4f53\u4efb\u52a1\u9700\u6c42\u548c\u8d44\u6e90\u9650\u5236&#xff0c;\u9700\u8981\u91c7\u7528\u4e0d\u540c\u7684PEFT\u914d\u7f6e\u7b56\u7565\u3002<\/p>\n<h4>4.1 \u901a\u7528\u573a\u666f&#xff08;\u5e73\u8861\u6027\u80fd\u4e0e\u6548\u7387&#xff09;<\/h4>\n<p>\u9002\u7528\u4e8e\u6587\u672c\u5206\u7c7b\u3001\u60c5\u611f\u5206\u6790\u7b49\u6807\u51c6\u4efb\u52a1&#xff1a;<\/p>\n<p>peft_config <span class=\"token operator\">&#061;<\/span> LoraConfig<span class=\"token punctuation\">(<\/span><br \/>\n    r<span class=\"token operator\">&#061;<\/span><span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u79e9<\/span><br \/>\n    lora_alpha<span class=\"token operator\">&#061;<\/span><span class=\"token number\">64<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># Alpha\u503c<\/span><br \/>\n    learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">2e<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u5b66\u4e60\u7387<\/span><br \/>\n    lora_dropout<span class=\"token operator\">&#061;<\/span><span class=\"token number\">0.05<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    target_modules<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;q_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;v_proj&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<h4>4.2 \u4f4e\u663e\u5b58\u573a\u666f&#xff08;8GB GPU&#xff09;<\/h4>\n<p>\u9488\u5bf9\u663e\u5b58\u6709\u9650\u7684\u6d88\u8d39\u7ea7GPU\u4f18\u5316&#xff1a;<\/p>\n<p>peft_config <span class=\"token operator\">&#061;<\/span> LoraConfig<span class=\"token punctuation\">(<\/span><br \/>\n    r<span class=\"token operator\">&#061;<\/span><span class=\"token number\">16<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u964d\u4f4e\u79e9<\/span><br \/>\n    lora_alpha<span class=\"token operator\">&#061;<\/span><span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u4fdd\u63012:1\u6bd4\u4f8b<\/span><br \/>\n    learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1e<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u964d\u4f4e\u5b66\u4e60\u7387\u4ee5\u63d0\u9ad8\u7a33\u5b9a\u6027<\/span><br \/>\n    load_in_8bit<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># 8\u4f4d\u91cf\u5316<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<h4>4.3 \u590d\u6742\u63a8\u7406\u573a\u666f<\/h4>\n<p>\u9488\u5bf9\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u7b49\u590d\u6742\u4efb\u52a1&#xff1a;<\/p>\n<p>peft_config <span class=\"token operator\">&#061;<\/span> LoraConfig<span class=\"token punctuation\">(<\/span><br \/>\n    r<span class=\"token operator\">&#061;<\/span><span class=\"token number\">64<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u590d\u6742\u4efb\u52a1\u9700\u8981\u9ad8\u79e9<\/span><br \/>\n    lora_alpha<span class=\"token operator\">&#061;<\/span><span class=\"token number\">128<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u7ef4\u63012:1\u6bd4\u4f8b<\/span><br \/>\n    learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">5e<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">5<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u964d\u4f4e\u5b66\u4e60\u7387\u9632\u6b62\u8fc7\u62df\u5408<\/span><br \/>\n    target_modules<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;q_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;v_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;k_proj&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;o_proj&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u5168\u6ce8\u610f\u529b\u5c42\u5fae\u8c03<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u8868&#xff1a;PEFT\u65b9\u6cd5\u9009\u62e9\u6307\u5357<\/p>\n<table>\n<tr>\u4efb\u52a1\u7c7b\u578b\u63a8\u8350\u65b9\u6cd5\u53c2\u6570\u914d\u7f6e\u5efa\u8bae\u8bad\u7ec3\u8d44\u6e90\u9700\u6c42<\/tr>\n<tbody>\n<tr>\n<td>\u6587\u672c\u5206\u7c7b\/\u60c5\u611f\u5206\u6790<\/td>\n<td>LoRA\u6216Prompt Tuning<\/td>\n<td>r&#061;16-32, alpha&#061;32-64<\/td>\n<td>\u4f4e&#xff08;8-16GB\u663e\u5b58&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u4ee3\u7801\u751f\u6210\/\u6570\u5b66\u63a8\u7406<\/td>\n<td>LoRA&#xff08;\u9ad8\u79e9&#xff09;\u6216AdaLoRA<\/td>\n<td>r&#061;32-64, alpha&#061;64-128<\/td>\n<td>\u4e2d&#xff08;16-24GB\u663e\u5b58&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u591a\u8f6e\u5bf9\u8bdd\u7cfb\u7edf<\/td>\n<td>Adapter\u6216Prefix Tuning<\/td>\n<td>\u9002\u914d\u5668\u74f6\u9888\u7ef4\u5ea6&#061;64-128<\/td>\n<td>\u4e2d\u9ad8&#xff08;24-40GB\u663e\u5b58&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u591a\u4efb\u52a1\u5b66\u4e60<\/td>\n<td>\u591a\u9002\u914d\u5668\u6216\u6df7\u5408\u65b9\u6cd5<\/td>\n<td>\u4efb\u52a1\u7279\u5b9a\u53c2\u6570\u914d\u7f6e<\/td>\n<td>\u53d6\u51b3\u4e8e\u6a21\u578b\u89c4\u6a21<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>5 \u5b9e\u8df5\u4f18\u5316\u4e0e\u6545\u969c\u6392\u9664<\/h3>\n<p>\u5728Ubuntu 22.04\u4e0a\u5b9e\u65bdPEFT\u5fae\u8c03\u65f6&#xff0c;\u9700\u8981\u6ce8\u610f\u4ee5\u4e0b\u4f18\u5316\u6280\u5de7\u548c\u5e38\u89c1\u95ee\u9898\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<h4>5.1 \u6027\u80fd\u4f18\u5316\u7b56\u7565<\/h4>\n<li>\n<p>\u68af\u5ea6\u68c0\u67e5\u70b9&#xff1a;\u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\u53ef\u4ee5\u5927\u5e45\u51cf\u5c11\u663e\u5b58\u5360\u7528&#xff0c;\u4f46\u4f1a\u589e\u52a0\u7ea620%\u7684\u8bad\u7ec3\u65f6\u95f4<\/p>\n<p>training_args <span class=\"token operator\">&#061;<\/span> TrainingArguments<span class=\"token punctuation\">(<\/span><br \/>\n    gradient_checkpointing<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token comment\"># &#8230; \u5176\u4ed6\u53c2\u6570<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<\/li>\n<li>\n<p>\u68af\u5ea6\u7d2f\u79ef&#xff1a;\u901a\u8fc7\u68af\u5ea6\u7d2f\u79ef\u6a21\u62df\u66f4\u5927\u7684\u6279\u5927\u5c0f<\/p>\n<p>training_args <span class=\"token operator\">&#061;<\/span> TrainingArguments<span class=\"token punctuation\">(<\/span><br \/>\n    per_device_train_batch_size<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    gradient_accumulation_steps<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u7b49\u6548\u6279\u5927\u5c0f&#061;8<\/span><br \/>\n    <span class=\"token comment\"># &#8230; \u5176\u4ed6\u53c2\u6570<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<\/li>\n<li>\n<p>\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3&#xff1a;\u4f7f\u7528FP16\u6216BF16\u7cbe\u5ea6\u52a0\u901f\u8bad\u7ec3<\/p>\n<p>training_args <span class=\"token operator\">&#061;<\/span> TrainingArguments<span class=\"token punctuation\">(<\/span><br \/>\n    fp16<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span>  <span class=\"token comment\"># \u6216 bf16&#061;True<\/span><br \/>\n    <span class=\"token comment\"># &#8230; \u5176\u4ed6\u53c2\u6570<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<\/li>\n<h4>5.2 \u5e38\u89c1\u95ee\u9898\u4e0e\u89e3\u51b3\u65b9\u6848<\/h4>\n<li>\n<p>\u663e\u5b58\u4e0d\u8db3&#xff1a;<\/p>\n<ul>\n<li>\u964d\u4f4eper_device_train_batch_size<\/li>\n<li>\u589e\u52a0gradient_accumulation_steps\u4fdd\u6301\u603b\u6279\u5927\u5c0f<\/li>\n<li>\u542f\u7528gradient_checkpointing<\/li>\n<li>\u4f7f\u7528load_in_8bit\u6216load_in_4bit\u91cf\u5316<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u8bad\u7ec3\u4e0d\u7a33\u5b9a&#xff1a;<\/p>\n<ul>\n<li>\u964d\u4f4e\u5b66\u4e60\u7387&#xff08;\u901a\u5e38\u51cf\u5c1150%&#xff09;<\/li>\n<li>\u589e\u52a0\u9884\u70ed\u6b65\u6570&#xff08;warmup steps&#xff09;<\/li>\n<li>\u8c03\u6574Alpha\u503c&#xff0c;\u4fdd\u6301Alpha\/Rank \u2265 1<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u8fc7\u62df\u5408&#xff1a;<\/p>\n<ul>\n<li>\u964d\u4f4e\u79e9&#xff08;\u51cf\u5c11\u6a21\u578b\u8868\u8fbe\u80fd\u529b&#xff09;<\/li>\n<li>\u589e\u52a0Dropout\u6bd4\u7387&#xff08;\u59820.1\u21920.3&#xff09;<\/li>\n<li>\u63d0\u524d\u505c\u6b62\u8bad\u7ec3<\/li>\n<\/ul>\n<\/li>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005&#xff1a;\u5434\u4e1a\u4eae<br \/>\n\u535a\u5ba2&#xff1a;wuyeliang.blog.csdn.net<br \/>\n1 PEFT\u6280\u672f\u6838\u5fc3\u539f\u7406<br \/>\n\u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;Parameter-Efficient Fine-Tuning, PEFT&#xff09;\u662f\u4e00\u79cd\u9488\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b&#xff08;\u5c24\u5176\u662f\u5927\u8bed\u8a00\u6a21\u578b&#xff09;\u7684\u5fae\u8c03\u7b56\u7565&#xff0c;\u5176\u6838\u5fc3\u601d\u60f3\u662f\u907f\u514d\u5bf9\u6a21\u578b\u7684\u5168\u90e8\u53c2\u6570\u8fdb\u884c\u66f4\u65b0&#xff0c;\u800c\u662f\u4ec5\u8c03\u6574\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u6216\u5f15\u5165\u5c11\u91cf\u989d\u5916\u53c2\u6570&#xff0c;\u4ece\u800c\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u548c\u5b58\u50a8\u6210\u672c&amp;#xf<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[5488,5721],"topic":[],"class_list":["post-83911","post","type-post","status-publish","format-standard","hentry","category-server","tag-peft","tag-5721"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/83911.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4f5c\u8005&#xff1a;\u5434\u4e1a\u4eae \u535a\u5ba2&#xff1a;wuyeliang.blog.csdn.net 1 PEFT\u6280\u672f\u6838\u5fc3\u539f\u7406 \u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;Parameter-Efficient Fine-Tuning, PEFT&#xff09;\u662f\u4e00\u79cd\u9488\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b&#xff08;\u5c24\u5176\u662f\u5927\u8bed\u8a00\u6a21\u578b&#xff09;\u7684\u5fae\u8c03\u7b56\u7565&#xff0c;\u5176\u6838\u5fc3\u601d\u60f3\u662f\u907f\u514d\u5bf9\u6a21\u578b\u7684\u5168\u90e8\u53c2\u6570\u8fdb\u884c\u66f4\u65b0&#xff0c;\u800c\u662f\u4ec5\u8c03\u6574\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u6216\u5f15\u5165\u5c11\u91cf\u989d\u5916\u53c2\u6570&#xff0c;\u4ece\u800c\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u548c\u5b58\u50a8\u6210\u672c&amp;#xf\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/83911.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-25T22:17:00+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/83911.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/83911.html\",\"name\":\"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-25T22:17:00+00:00\",\"dateModified\":\"2026-07-25T22:17:00+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/83911.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/83911.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/83911.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/83911.html","og_locale":"zh_CN","og_type":"article","og_title":"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4f5c\u8005&#xff1a;\u5434\u4e1a\u4eae \u535a\u5ba2&#xff1a;wuyeliang.blog.csdn.net 1 PEFT\u6280\u672f\u6838\u5fc3\u539f\u7406 \u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;Parameter-Efficient Fine-Tuning, PEFT&#xff09;\u662f\u4e00\u79cd\u9488\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b&#xff08;\u5c24\u5176\u662f\u5927\u8bed\u8a00\u6a21\u578b&#xff09;\u7684\u5fae\u8c03\u7b56\u7565&#xff0c;\u5176\u6838\u5fc3\u601d\u60f3\u662f\u907f\u514d\u5bf9\u6a21\u578b\u7684\u5168\u90e8\u53c2\u6570\u8fdb\u884c\u66f4\u65b0&#xff0c;\u800c\u662f\u4ec5\u8c03\u6574\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u6216\u5f15\u5165\u5c11\u91cf\u989d\u5916\u53c2\u6570&#xff0c;\u4ece\u800c\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u548c\u5b58\u50a8\u6210\u672c&amp;#xf","og_url":"https:\/\/www.wsisp.com\/helps\/83911.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-25T22:17:00+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/83911.html","url":"https:\/\/www.wsisp.com\/helps\/83911.html","name":"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-25T22:17:00+00:00","dateModified":"2026-07-25T22:17:00+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/83911.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/83911.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/83911.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"PEFT 2.0\u8fdb\u9636\uff1aUbuntu\u670d\u52a1\u5668\u4e0a\u7684\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\u4e0e\u4f18\u5316"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/83911","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=83911"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/83911\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=83911"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=83911"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=83911"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=83911"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}