{"id":169,"date":"2026-04-27T16:59:09","date_gmt":"2026-04-27T08:59:09","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=169"},"modified":"2026-04-27T16:59:09","modified_gmt":"2026-04-27T08:59:09","slug":"peft-deep-dive-parameter-efficient-fine-tuning","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/27\/peft-deep-dive-parameter-efficient-fine-tuning\/","title":{"rendered":"PEFT \u6df1\u5ea6\u89e3\u6790\uff1a\u5982\u4f55\u7528\u51e0MB\u7684\u53c2\u6570\u5fae\u8c03\u767e\u4ebf\u7ea7\u5927\u6a21\u578b"},"content":{"rendered":"<h2>\u4e3a\u4ec0\u4e48\u5168\u53c2\u6570\u5fae\u8c03\u5728\u5927\u6a21\u578b\u65f6\u4ee3\u884c\u4e0d\u901a\uff1f<\/h2>\n<p>\u5f53\u4f60\u9762\u5bf9\u4e00\u4e2a100\u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff0c\u6bd4\u5982 <code>bigscience\/mt0-xxl<\/code>\uff0c\u4f20\u7edf\u505a\u6cd5\u662f\u52a0\u8f7d\u6574\u4e2a\u6a21\u578b\uff0c\u5bf9\u6240\u6709\u53c2\u6570\u8fdb\u884c\u68af\u5ea6\u66f4\u65b0\u3002\u8fd9\u542c\u8d77\u6765\u5f88\u76f4\u63a5\uff0c\u4f46\u4ee3\u4ef7\u60ca\u4eba\uff1a<\/p>\n<ul>\n<li><strong>\u663e\u5b58\u5360\u7528<\/strong>\uff1a\u5355\u6b21\u524d\u5411+\u53cd\u5411\u4f20\u64ad\u9700\u8981\u8d85\u8fc780GB\u663e\u5b58\uff0c\u8fdc\u8d85\u6d88\u8d39\u7ea7\u663e\u5361\uff08\u5982RTX 3090\u768424GB\uff09\u3002<\/li>\n<li><strong>\u5b58\u50a8\u6210\u672c<\/strong>\uff1a\u6bcf\u4e2a\u4e0b\u6e38\u4efb\u52a1\u90fd\u8981\u4fdd\u5b58\u4e00\u4e2a\u72ec\u7acb\u768440GB\u6a21\u578b\u526f\u672c\uff0c10\u4e2a\u4efb\u52a1\u5c31\u662f400GB\u3002<\/li>\n<li><strong>\u707e\u96be\u6027\u9057\u5fd8<\/strong>\uff1a\u5168\u53c2\u6570\u66f4\u65b0\u4f1a\u8986\u76d6\u539f\u59cb\u9884\u8bad\u7ec3\u77e5\u8bc6\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u65b0\u4efb\u52a1\u4e0a\u8868\u73b0\u597d\uff0c\u5374\u5fd8\u4e86\u539f\u6765\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u95ee\u9898\u7684\u672c\u8d28\u662f\uff1a<strong>\u6211\u4eec\u771f\u7684\u9700\u8981\u66f4\u65b0\u6bcf\u4e00\u4e2a\u53c2\u6570\u5417\uff1f<\/strong><\/p>\n<h2>PEFT \u7684\u6838\u5fc3\u601d\u60f3\uff1a\u4e0d\u6539\u5927\u53a6\uff0c\u53ea\u52a0\u697c\u68af<\/h2>\n<p>PEFT \u7684\u54f2\u5b66\u662f\uff1a<strong>\u51bb\u7ed3\u4e3b\u5e72\uff0c\u53ea\u5fae\u8c03\u6781\u5c0f\u7684\u201c\u9002\u914d\u5c42\u201d<\/strong>\u3002<\/p>\n<p>\u60f3\u8c61\u4f60\u6709\u4e00\u680b\u767e\u5c42\u6469\u5929\u5927\u697c\uff08\u9884\u8bad\u7ec3\u6a21\u578b\uff09\uff0c\u4f60\u60f3\u8ba9\u5b83\u9002\u5e94\u4e00\u4e2a\u65b0\u7528\u9014\uff08\u6bd4\u5982\u533b\u7597\u95ee\u7b54\uff09\u3002\u4f20\u7edf\u505a\u6cd5\u662f\u628a\u6574\u680b\u697c\u62c6\u4e86\u91cd\u76d6\uff08\u5168\u53c2\u6570\u5fae\u8c03\uff09\u3002PEFT\u7684\u505a\u6cd5\u662f\uff1a\u5728\u697c\u9876\u52a0\u4e00\u4e2a\u8f7b\u5de7\u7684\u65cb\u8f6c\u697c\u68af\uff08\u9002\u914d\u6a21\u5757\uff09\uff0c\u8ba9\u8bbf\u5ba2\u80fd\u76f4\u8fbe\u65b0\u697c\u5c42\uff0c\u800c\u5927\u697c\u4e3b\u4f53\u7eb9\u4e1d\u4e0d\u52a8\u3002<\/p>\n<p>\u8fd9\u4e2a\u201c\u65cb\u8f6c\u697c\u68af\u201d\u5c31\u662f\u53ef\u8bad\u7ec3\u7684<strong>\u4f4e\u79e9\u77e9\u9635<\/strong>\uff0c\u5b83\u53ea\u6709\u51e0\u5343\u5230\u51e0\u4e07\u4e2a\u53c2\u6570\uff0c\u5374\u80fd\u5f15\u5bfc\u6574\u4e2a\u5927\u6a21\u578b\u7684\u8f93\u51fa\u65b9\u5411\u3002<\/p>\n<h2>LoRA\uff1a\u4f4e\u79e9\u9002\u5e94\u7684\u6570\u5b66\u9b54\u672f<\/h2>\n<p>LoRA\uff08Low-Rank Adaptation\uff09\u662fPEFT\u4e2d\u6700\u6210\u529f\u7684\u65b9\u6848\u3002\u5b83\u7684\u6838\u5fc3\u662f\u77e9\u9635\u5206\u89e3\uff1a<\/p>\n<p>\u5728Transformer\u7684\u6ce8\u610f\u529b\u673a\u5236\u4e2d\uff0c\u6743\u91cd\u77e9\u9635 $ W \\in \\mathbb{R}^{d \\times k} $ \u901a\u5e38\u5f88\u5927\u3002LoRA\u4e0d\u76f4\u63a5\u4fee\u6539 $ W $\uff0c\u800c\u662f\u5f15\u5165\u4e24\u4e2a\u5c0f\u77e9\u9635\uff1a<\/p>\n<p>$$ \\Delta W = B \\cdot A, \\quad A \\in \\mathbb{R}^{d \\times r}, B \\in \\mathbb{R}^{r \\times k} $$<\/p>\n<p>\u5176\u4e2d $ r \\ll \\min(d, k) $\uff0c\u901a\u5e38\u53d68\u621616\u3002<\/p>\n<ul>\n<li>\u539f\u59cb\u6743\u91cd $ W $\uff1a\u51bb\u7ed3\uff0c\u4e0d\u66f4\u65b0\u3002<\/li>\n<li>\u65b0\u589e\u53c2\u6570\uff1a$ A $ \u548c $ B $\uff0c\u5171 $ r \\cdot (d + k) $ \u4e2a\uff0c\u8fdc\u5c0f\u4e8e $ d \\cdot k $\u3002<\/li>\n<li>\u63a8\u7406\u65f6\uff1a$ W_{\\text{new}} = W + B \\cdot A $\uff0c\u7b49\u4ef7\u4e8e\u4e00\u4e2a\u5b8c\u6574\u6743\u91cd\uff0c\u4f46\u8bad\u7ec3\u65f6\u53ea\u66f4\u65b0A\u3001B\u3002<\/li>\n<\/ul>\n<p><strong>\u4e3a\u4ec0\u4e48\u4f4e\u79e9\u6709\u6548\uff1f<\/strong><\/p>\n<p>\u795e\u7ecf\u7f51\u7edc\u7684\u53c2\u6570\u7a7a\u95f4\u5b58\u5728\u201c\u4f4e\u79e9\u7ed3\u6784\u201d\u2014\u2014\u5927\u91cf\u5197\u4f59\u4fe1\u606f\u96c6\u4e2d\u5728\u5c11\u6570\u4e3b\u6210\u5206\u4e0a\u3002LoRA\u901a\u8fc7\u4f4e\u79e9\u5206\u89e3\uff0c\u6070\u597d\u6355\u6349\u4e86\u8fd9\u4e9b\u5173\u952e\u65b9\u5411\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u5373\u4f7f $ r=8 $\uff0c\u4e5f\u80fd\u903c\u8fd1\u5168\u53c2\u6570\u5fae\u8c03\u7684\u6548\u679c\u3002<\/p>\n<h2>\u5b9e\u9645\u8bad\u7ec3\uff1a\u51e0\u884c\u4ee3\u7801\u7684\u9769\u547d<\/h2>\n<pre><code class=\"language-python\">from transformers import AutoModelForSeq2SeqLM\nfrom peft import get_peft_model, LoraConfig, TaskType\n\nmodel = AutoModelForSeq2SeqLM.from_pretrained(&quot;bigscience\/mt0-large&quot;)\n\npeft_config = LoraConfig(\n    task_type=TaskType.SEQ_2_SEQ_LM,\n    r=8,           # \u4f4e\u79e9\u7ef4\u5ea6\n    lora_alpha=32, # \u7f29\u653e\u56e0\u5b50\n    lora_dropout=0.1\n)\n\nmodel = get_peft_model(model, peft_config)\nmodel.print_trainable_parameters()\n# \u8f93\u51fa\uff1a\u53ef\u8bad\u7ec3\u53c2\u6570\uff1a235\u4e07 || \u603b\u53c2\u6570\uff1a12.3\u4ebf || \u53ef\u8bad\u7ec3\u6bd4\u4f8b\uff1a0.19%<\/code><\/pre>\n<p>\u4f60\u53ea\u8bad\u7ec3\u4e860.19%\u7684\u53c2\u6570\uff01\u4f46\u6a21\u578b\u6027\u80fd\u51e0\u4e4e\u4e0e\u5168\u5fae\u8c03\u6301\u5e73\u3002<\/p>\n<p>\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c<code>model.save_pretrained()<\/code> \u53ea\u4fdd\u5b58\u4e24\u4e2a\u6587\u4ef6\uff1a<\/p>\n<ul>\n<li><code>adapter_config.json<\/code>\uff1aLoRA\u914d\u7f6e<\/li>\n<li><code>adapter_model.bin<\/code>\uff1a\u4ec519MB\u7684A\u3001B\u77e9\u9635<\/li>\n<\/ul>\n<p>\u63a8\u7406\u65f6\uff0c\u52a0\u8f7d\u4e3b\u6a21\u578b + \u52a0\u8f7d\u9002\u914d\u5668\uff0c\u5373\u53ef\u590d\u7528\uff1a<\/p>\n<pre><code class=\"language-python\">model = AutoModelForSeq2SeqLM.from_pretrained(&quot;base-model&quot;)\nmodel = PeftModel.from_pretrained(model, &quot;adapter-path&quot;)<\/code><\/pre>\n<h2>\u4e3a\u4ec0\u4e48\u8fd9\u80fd\u89e3\u51b3\u707e\u96be\u6027\u9057\u5fd8\uff1f<\/h2>\n<p>\u56e0\u4e3a\u4e3b\u6a21\u578b\u53c2\u6570\u5b8c\u5168\u51bb\u7ed3\uff0c\u539f\u59cb\u77e5\u8bc6\u88ab\u201c\u9501\u5b9a\u201d\u5728\u6743\u91cd\u4e2d\u3002LoRA\u6a21\u5757\u50cf\u4e00\u4e2a\u201c\u63d0\u793a\u63a7\u5236\u5668\u201d\uff0c\u53ea\u5728\u8f93\u51fa\u5c42\u65bd\u52a0\u5fae\u5c0f\u6270\u52a8\uff0c\u5f15\u5bfc\u6a21\u578b\u8c03\u7528\u5df2\u6709\u77e5\u8bc6\u53bb\u89e3\u51b3\u65b0\u4efb\u52a1\uff0c\u800c\u975e\u91cd\u5199\u8bb0\u5fc6\u3002<\/p>\n<p>\u8fd9\u4f7f\u5f97\u540c\u4e00\u4e2a\u57fa\u7840\u6a21\u578b\uff0c\u53ef\u4ee5\u540c\u65f6\u52a0\u8f7d\u591a\u4e2aLoRA\u9002\u914d\u5668\uff0c\u5207\u6362\u4efb\u52a1\u5982\u540c\u6362\u63d2\u5934\uff1a<\/p>\n<ul>\n<li>\u533b\u7597\u95ee\u7b54 \u2192 \u52a0\u8f7d medical_lora.bin<\/li>\n<li>\u6cd5\u5f8b\u6587\u4e66 \u2192 \u52a0\u8f7d legal_lora.bin<\/li>\n<li>\u4ee3\u7801\u751f\u6210 \u2192 \u52a0\u8f7d code_lora.bin<\/li>\n<\/ul>\n<p>\u6240\u6709\u9002\u914d\u5668\u5171\u7528\u540c\u4e00\u4e2a40GB\u4e3b\u6a21\u578b\uff0c\u603b\u5b58\u50a8\u4ec5\u589e\u52a0\u51e0\u5341MB\u3002<\/p>\n<h2>\u9002\u7528\u573a\u666f\u4e0d\u6b62\u4e8e\u6587\u672c<\/h2>\n<p>PEFT\u7684\u5a01\u529b\u4e0d\u9650\u4e8eNLP\u3002\u5728Stable Diffusion\u4e2d\uff0cDreamBooth\u901a\u8fc7LoRA\u5fae\u8c03\u56fe\u50cf\u751f\u6210\u5668\uff0c\u4ec5\u752810\u5f20\u56fe\u5c31\u80fd\u8ba9\u6a21\u578b\u5b66\u4f1a\u753b\u4f60\u7684\u5ba0\u7269\u72d7\uff0c\u800c\u6a21\u578b\u4f53\u79ef\u4ece10GB\u538b\u7f29\u5230100MB\u4ee5\u5185\u3002<\/p>\n<p>\u5728\u8bed\u97f3\u8bc6\u522b\uff08Whisper\uff09\u3001\u89c6\u89c9\uff08ViT\uff09\u4e2d\uff0cPEFT\u540c\u6837\u5927\u5e45\u964d\u4f4e\u8bad\u7ec3\u95e8\u69db\u3002<\/p>\n<h2>\u603b\u7ed3\uff1aPEFT\u4e0d\u662f\u6280\u5de7\uff0c\u662f\u8303\u5f0f\u8f6c\u79fb<\/h2>\n<p>PEFT\u7684\u672c\u8d28\u662f<strong>\u4ece\u201c\u5168\u6a21\u578b\u5fae\u8c03\u201d\u8f6c\u5411\u201c\u6a21\u578b+\u63d2\u4ef6\u201d\u67b6\u6784<\/strong>\u3002\u5b83\u8ba9\u5927\u6a21\u578b\u4ece\u201c\u5962\u4f88\u54c1\u201d\u53d8\u6210\u201c\u53ef\u63d2\u62d4\u7684\u57fa\u7840\u8bbe\u65bd\u201d\u3002<\/p>\n<ul>\n<li>\u2705 \u6d88\u8d39\u7ea7\u786c\u4ef6\u53ef\u8bad\u7ec3\u767e\u4ebf\u6a21\u578b<\/li>\n<li>\u2705 \u591a\u4efb\u52a1\u5171\u4eab\u4e00\u4e2a\u4e3b\u6a21\u578b<\/li>\n<li>\u2705 \u5b58\u50a8\u6210\u672c\u4e0b\u964d1000\u500d<\/li>\n<li>\u2705 \u907f\u514d\u707e\u96be\u6027\u9057\u5fd8<\/li>\n<li>\u2705 \u652f\u6301\u591a\u6a21\u6001<\/li>\n<\/ul>\n<p>\u672a\u6765\uff0c\u6211\u4eec\u53ef\u80fd\u4e0d\u518d\u4e0b\u8f7d\u201c\u5fae\u8c03\u540e\u7684\u6a21\u578b\u201d\uff0c\u800c\u662f\u4e0b\u8f7d\u201c\u6a21\u578b+\u9002\u914d\u5668\u201d\u3002PEFT\uff0c\u6b63\u5728\u91cd\u65b0\u5b9a\u4e49AI\u7684\u90e8\u7f72\u65b9\u5f0f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.bojinhu.xyz\/wp-content\/uploads\/2026\/04\/cover-1.png\" alt=\"Cover\" \/> <img decoding=\"async\" src=\"https:\/\/www.bojinhu.xyz\/wp-content\/uploads\/2026\/04\/lora-matrix.png\" alt=\"Lora Matrix\" \/> <img decoding=\"async\" src=\"https:\/\/www.bojinhu.xyz\/wp-content\/uploads\/2026\/04\/adapter-loading.png\" alt=\"Adapter Loading\" \/><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u6df1\u5165\u5256\u6790PEFT\uff08\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff09\u7684\u6838\u5fc3\u673a\u5236\uff0c\u89e3\u91caLoRA\u5982\u4f55\u901a\u8fc7\u4f4e\u79e9\u5206\u89e3\u5728\u51bb\u7ed3\u4e3b\u6a21\u578b\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u5ab2\u7f8e\u5168\u53c2\u6570\u5fae\u8c03\u7684\u6548\u679c\uff0c\u63ed\u793a\u5176\u5728\u6d88\u8d39\u7ea7\u786c\u4ef6\u4e0a\u8bad\u7ec3\u767e\u4ebf\u6a21\u578b\u7684\u5e95\u5c42\u539f\u7406\u3002<\/p>\n","protected":false},"author":2,"featured_media":166,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[68,66,64,67,65],"class_list":["post-169","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-hugging-face","tag-lora","tag-peft","tag-67","tag-65"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/169","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=169"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions"}],"predecessor-version":[{"id":170,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions\/170"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/166"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=169"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=169"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=169"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}